# Asymmetry Computing — full site text This is the complete text of the main pages of https://asymmetrycomputing.com, flattened to markdown for retrieval systems and agents that prefer one plain text source over parsing HTML. A shorter structured index is at https://asymmetrycomputing.com/llms.txt Asymmetry Computing builds PRISM, a real-time optimization engine for institutional finance that returns a feasible, near-optimal, fully auditable answer inside a hard deadline. Founded 2025 by Debdoot Ghosh. "PRISM" is a heavily reused product name across software. This engine is "Asymmetry PRISM" or "PRISM by Asymmetry Computing" — it is unrelated to the Oxford probabilistic model checker, GraphPad Prism, the PrismLibrary XAML framework, or similarly named fintech products. Primary sources: - Paper: https://arxiv.org/abs/2606.23367 (arXiv:2606.23367) - Public evaluation code: https://github.com/AsymmetryComputing/prism-public-evaluation - GitHub organization: https://github.com/AsymmetryComputing - LinkedIn: https://www.linkedin.com/company/asymmetrycomputing - Contact: debdoot@asymmetrycomputing.com --- # Home Source: https://asymmetrycomputing.com/ PRISM · One real-time optimization engine ## Extreme scale. Hyper-personalized. In real time. Fully deterministic · Provably feasible · Auditable The optimization engine for problems too big and too personal for conventional solvers — solved orders of magnitude faster, the instant the world moves. One engine, proven in finance and expanding across real-time operations. Start an 8-week paid pilot → See the evidence 8-week paid pilot on your data — you get a benchmark report on your real problem and reference pricing for production. Prefer a form? Request a pilot → Direct Indexing $238K Portfolio Decisioning <1 s Energy Dispatch ~5 ms Reproduce the proof FF30 Real-time solve Live solve map Real-Time Energy Dispatch Battery & DER dispatch · within the grid control-tick Deadline ~5 ms Feasible 100% Changed few Data CAISO DISPATCH: feasible, audited plan on the wire before the tick closes Start here ### What this actually does, in plain English. No background assumed. Every specialist term on this site is defined on the glossary page . Some decisions have three awkward properties at the same time. There are far too many possible answers to check by hand. The answer has to obey strict rules. And it has to arrive before a particular moment, or it is worthless — however good it would have been. Two real examples. A wealth manager runs fifty thousand individual investment accounts. Every night, before the market opens, each one needs its own list of what to buy and sell — respecting that client's tax position, restrictions and risk limits. Miss the open and the trades don't happen. Separately, an electricity grid operator has thousands of batteries spread across a region. Every few seconds, each one needs a new instruction as power prices move. The grid does not wait; miss the moment and the opportunity is gone. These two look nothing alike. Underneath they are the same problem: search an enormous number of options, obey every rule, and finish on time. PRISM is the software that does that. It returns an answer that breaks none of the rules, is as close to the best available as makes any practical difference, arrives inside the deadline, and comes with a record of how it was reached — so an auditor or regulator can check the decision later. Run it twice on the same inputs and you get exactly the same answer. The honest short version. Existing tools already solve these problems well at small scale. They struggle when the number of accounts, assets or devices gets large and a hard deadline applies. That gap is the only thing we sell into — and where we don't win, we publish that too . Where PRISM is proven ### Anchored in finance. One engine, expanding. The deepest, money-grade proof is in finance — direct indexing and real-time portfolio decisioning, on real data, losing cases kept in. The same core extends to any high-stakes decision that must be re-solved correctly, on time, the instant the world moves. Finance · Direct Indexing · Flagship #### Direct Indexing & Tax-Aware Householding Personalized, tax-aware portfolios across thousands of accounts — deterministic, auditable, and fast at fleet scale, with the tail reported to the percentile. The deepest evidence base on the site. $238K Full harvestable tax budget captured on a $5M, 192-name book; a 500,000-account book priced in ~2 min on one GPU (~25–30× a tuned baseline), p50/p95/p99 reported. Explore Direct Indexing → Finance · Real-Time Risk #### Real-Time Portfolio Decisioning Optimal rebalancing plus full risk attribution inside a hard latency budget — at universe sizes where standard optimizers stall. <1 s A 75,000-asset rebalance with factor/specific risk attribution and an audit trail, returned in under a second. Explore Portfolio Decisioning → Energy · Grid · DER/VPP #### Real-Time Energy Dispatch Re-optimize battery & distributed-energy dispatch the instant prices or forecasts move — within the grid control-tick, changing only the few intervals worth changing. ~5 ms Feasible, audited dispatch delivered within deadlines from 500 ms down to ~5 ms — demonstrated on real California ISO data. Explore Energy Dispatch → Beyond the three above, the same engine is expanding across real-time operations — telecom & satellite resource allocation, AI/HPC cluster scheduling, and pricing & budget allocation — with molecular candidate selection on the research roadmap. Earlier-stage work is under NDA. Selected measured results ### Measured on real data. Losing cases kept in. Every figure here traces to a recorded run on real data — with losing cases kept in, not cherry-picked. The peer-reviewable write-up is arXiv:2606.23367 ; the evaluation artifacts are public on GitHub ; the provenance of each number is on the research page . Fleet throughput 10.2× more accounts priced per core-second 100,000 accounts priced in ~11.6 min · deterministic & auditable Tax alpha captured $238K harvestable tax budget on a $5M, 192-name book full budget band vs commercial baseline on the structured lane Real-time dispatch ~5 ms feasible, audited grid dispatch plan delivered within deadlines from 500 ms down to ~5 ms · real CAISO data 100,004-asset structured solve. A real-data structured solve across 100,004 assets and 47+ exchanges, completed deterministically with a full audit trail. The current recorded solve time on the 100k structured lane is 2.16 s . Reproduce it yourself. On public Kenneth French (FF30) data — no licensed data, nothing to take on faith — PRISM runs 23.28× faster than a commercial baseline at matched quality. The fastest way to trust a benchmark is to run it on data you already have. See the methodology → Measured on real hardware Every figure on this page traces to a recorded run on a dedicated RTX 4000 Ada GPU — real data, losing cases kept in , not cherry-picked. Why PRISM ### Built for operational reality. The answer arrives inside the deadline that matters — fast where scale demands it, feasible and auditable every time. One engine, the same contract across every domain. #### Speed Built for the books where scale is the problem — hundreds of thousands of accounts, or universes in the tens of thousands of names, answered inside the window you actually operate in. Per-lane speed figures are being re-measured under a contention guard before republication. #### Reproducible Audit-ready delivery with explicit, documented execution paths. Deterministic, content-hashed, re-derivable — every trade traces to your constraints. #### Works with your stack HTTPS API. Slots into OMS/EMS, grid control loops, custodian APIs, and risk models. Deploy in your VPC, on-prem, or air-gapped — no per-seat solver license. #### Proven at scale 75,257 real assets on the proof lane and a 100,004-asset structured suite. Evidence spans head-to-head, routed scale, and backtest surfaces. Constraints All constraints enforced simultaneously. Deterministic solve. Auditable. Hard real-time deadlines (control-tick to overnight) Ramp limits and state-of-charge (battery / DER) Wash-sale handling, lot-level tax accounting Tracking error and risk-factor caps Turnover limits and transaction-cost penalties Liquidity constraints (ADV, min trade sizes) Position-count and cardinality limits Sector and ESG exposure limits The gap conventional solvers can't close ### The answer that arrives too late is worth nothing. Across grids, trading desks, and personalized books, the world moves on a clock you don't control. Exact solvers find a beautiful answer — long after the moment to act has passed. Simpler methods keep up but leave value, or feasibility, on the table. PRISM lives in that gap: a feasible, near-optimal, fully auditable answer inside the deadline — the same engine, whether the window is a grid control-tick or an overnight book. Read the benchmarks → The contract, on every solve Feasible every constraint honoured Auditable every leg reported Ingest, per-account build and solve are broken out separately on every result — no hidden legs, losing cases kept in. Speed figures are being re-measured under a contention guard and will be republished per lane with the load recorded alongside. Built for the desk Deterministic & content-hashed Fully reproducible audit trail VPC, on-prem, or air-gapped No per-seat solver license ### Bring your hardest real-time problem. We'll prove it on your data. 8-week paid pilot on your data — you get a benchmark report on your real problem and reference pricing for production. You set the pass/fail metric before we start; every losing case is shown. Start an 8-week paid pilot → Prefer a form? Request a pilot → --- # Applications Source: https://asymmetrycomputing.com/applications Applications ## Real-time optimization, everywhere. PRISM's optimization core returns a feasible, near-optimal, fully auditable answer inside a hard real-time deadline — at a scale and speed where conventional solvers can't respond in time. The same engine powers new domains: In plain English. Every application below is the same underlying problem: choose the best option from an enormous number of possibilities, obey every rule, and finish before a deadline. Only the subject matter changes — investment accounts, batteries on a power grid, a bank's capital. Unfamiliar words? See the glossary . Start an 8-week paid pilot → See the evidence ~5 ms Feasible, audited grid dispatch delivered within deadlines this tight 75 K Real-asset universe rebalanced — with full risk attribution — in well under a second 10 K Accounts per fleet batch, with p50 / p95 / p99 latency reporting 1 engine One optimization core behind every domain on this page One engine, many domains The same core optimizes the trading book and the power grid — the two worlds in one frame. Where PRISM is proven ### Anchored in finance. One engine, expanding. The deepest evidence is in finance — direct indexing and real-time portfolio decisioning. The same core extends to any high-stakes decision that must be re-solved correctly, on time, the instant the world moves. Finance · Direct Indexing · Flagship #### Scaled Direct Indexing & Tax-Aware Householding Personalized, tax-aware portfolios across thousands of accounts — deterministic, auditable, and fast at fleet scale. PRISM's deepest evidence base. Each client owns real shares in their own account, so each one needs its own buy/sell list every night — tailored to their tax position. This is doing that for tens of thousands of accounts before the market opens. Demonstrated Production-grade fleet engine: 100 / 1,000 / 10,000-account batches with full latency reporting (p50/p95/p99), tax-lot handling, and a full audit trail. Explore Direct Indexing at scale → Fleet batch · per-account, at scale Batch Latency ladder is illustrative · p50 / p95 / p99 reported for every batch 100 accounts 1,000 10,000 Finance · Real-Time Risk #### Real-Time Portfolio Decisioning Optimal rebalancing plus full risk attribution inside a hard latency budget — at universe sizes where standard optimizers stall. A big portfolio drifts out of shape as prices move. This works out what to buy and sell to bring it back in line, fast enough to still act on, and shows exactly where its risk is coming from. Demonstrated Demonstrated on a real 75,000-asset dataset: an optimal rebalance, a factor/specific risk breakdown, and an audit trail returned in well under a second — across tens of thousands of holdings. Explore Portfolio Decisioning → Rebalance + risk attribution · output < 1s Factor / specific split shown is illustrative · full attribution returned per run 75,000 assets Returned in < 1 second Energy · Grid · DER/VPP #### Real-Time Energy Dispatch Re-optimize battery & distributed-energy dispatch the instant prices or forecasts move — within the grid control-tick. Thousands of batteries spread across an electricity grid each need a fresh instruction every few seconds as power prices move — and each instruction has to be something the equipment can physically do. Demonstrated Demonstrated on real California ISO grid data: a feasible, audited dispatch plan delivered within deadlines from 500 ms down to ~5 ms, changing only the few intervals worth changing. Explore Energy Dispatch → Dispatch · re-optimized intervals Live tick Only the few intervals worth changing move · the rest hold Delivered within 500 ms → ~5 ms ✓ Feasible · audited Demonstrated results on the datasets described — not guarantees. Comparators are labeled generically as conventional / standard solvers. Schematic elements in the panels above are marked illustrative. Try the interactive models: tax-loss harvesting · rebalance latency · FRTB capital · energy dispatch . How it fits ### One engine. One clean contract. Inputs in; a feasible, near-optimal, fully auditable answer out — inside your deadline. The methods are proprietary; the interface is the same across every domain above. Inputs - Live prices & forecasts - Positions & constraints - Your risk view - Cost & hard deadline → PRISM Real-time optimization core One engine, a hard deadline, deterministic. → Outputs - Feasible, near-optimal answer - Only the changes worth making - Full p50 / p95 / p99 latency - Content-hashed audit log #### Delivered in time The answer arrives inside the deadline that matters — a grid control-tick, a latency budget, an overnight window — not whenever the solver happens to finish. #### Auditable by design Every decision traces to your inputs and constraints, with a content-hashed, reproducible record — the same audit trail across energy, finance, and wealth. ### One engine. Many real-time, high-stakes decisions — delivered in time, every time, with a full audit trail. 8-week paid pilot on your data — you get a benchmark report on your real problem and reference pricing for production. Your data, your constraints, your deadline; every losing case shown. Start an 8-week paid pilot → Prefer a form? Request a pilot → --- # Real-Time Energy Dispatch Source: https://asymmetrycomputing.com/applications/energy-dispatch Application · Energy Dispatch ## Re-optimize dispatch within the control-tick. Re-optimize battery & distributed-energy dispatch the instant prices or forecasts move — within the grid control-tick. The price signal and the forecast change continuously; a dispatch plan that arrives after the tick is already stale. PRISM returns a feasible, audited plan inside the deadline, changing only the few intervals worth changing. In plain English. An electricity grid has to match supply and demand continuously. Batteries and other small power resources spread across the network need fresh instructions every few seconds as prices and demand move — and those instructions must be physically possible for the equipment to follow. This page is about computing them inside the few milliseconds available. Unfamiliar words? See the glossary . ▶ Run the live demo Start an 8-week paid pilot → No account · runs in your browser · real NYISO data · live now ↓ ~5 ms Tightest deadline a feasible, audited plan was delivered within 500 ms→5ms Range of hard deadlines met on real California ISO grid data few Intervals changed — only the ones worth changing move 100 % Feasible & audited — every plan, every tick PRISM · VPP / DER Dispatch Every battery, EV and solar inverter on this grid is a decision variable — re-optimised together, inside the 5-minute market tick . What it does ### One engine that runs your whole fleet — in plain terms. A virtual power plant is thousands of batteries, EVs and solar inverters owned by different people. PRISM operates them together, as one power plant, on the market's clock. 01 #### Coordinates the whole fleet Decides — for every device, every five minutes — whether it charges, discharges or holds, to capture the most value from moving prices. 02 #### Respects the real grid Honours every physical limit: feeder capacity, ramp rates, battery state-of-charge and efficiency. The plan is one you can actually run. 03 #### Beats the clock Returns a feasible, audited plan inside the 5-minute market window — at fleet sizes where ordinary solvers run out of memory. ### The tick doesn't wait. Neither can the plan. Grid prices and forecasts move on a clock you don't control. The job isn't to find a beautiful schedule eventually — it's to return a feasible, audited dispatch the moment the inputs change, inside the control-tick, and to disturb only what genuinely needs to move. Dispatch · re-optimized intervals Live tick Only the few intervals worth changing move · the rest hold Delivered within 500 ms → ~5 ms ✓ Feasible · audited 01 · The problem ### A dispatch plan has a shelf life measured in milliseconds. Batteries, distributed energy resources, and virtual power plants live inside a market and a physical grid that never hold still. Prices update, forecasts revise, telemetry arrives — and each change can make the plan you're executing the wrong one. The window to respond is a control-tick, not a coffee break. Miss it and you're dispatching against conditions that no longer exist. That turns dispatch into a hard real-time problem. The value isn't only in the quality of the schedule — it's in whether a feasible, defensible answer is on the wire before the tick closes. Conventional solvers can produce a good plan given enough time; the constraint here is that there isn't enough time, repeatedly, all day. Why coordination pays Every home on this street shares one feeder back to the grid. That shared limit is exactly what makes coordinating the fleet worth 12–28% more. 02 · Why it's hard ### Feasible, on a deadline, without churning the whole plan. Each of these is manageable alone. Together, on a millisecond clock, they're why a plan that's merely "optimal eventually" doesn't help. #### The deadline is the spec An answer after the control-tick is the wrong answer. The plan has to be feasible and on the wire inside the window — every tick, not on average. #### Intervals are coupled State of charge, ramp limits, and commitments link the intervals together, so you can't re-solve one slot in isolation — the plan moves as a whole. #### Stability matters Re-writing the entire schedule every tick is operationally noisy. The plan should change only the few intervals genuinely worth changing. 03 · How PRISM fits ### A black box with a clean contract. You bring live inputs and a hard deadline; PRISM returns a feasible, audited dispatch plan inside it. The methods are proprietary; the interface is simple. Inputs - Live prices & forecasts - Battery / DER state - Ramp & SoC constraints - Hard tick deadline → PRISM Real-time optimization core One engine, a hard deadline, deterministic. → Outputs - Feasible dispatch plan - Only the intervals worth changing - Delivered within the tick - Content-hashed audit log 04 · The evidence ### Demonstrated on real California ISO grid data. Demonstrated results on the dataset described — not a guarantee. Comparators are referred to generically as conventional / standard solvers. Deadline met ~5 ms A feasible, audited dispatch plan delivered within deadlines from 500 ms down to ~5 ms — across the real-time range a grid control loop actually runs at. Change discipline few The plan changed only the few intervals worth changing, rather than re-writing the whole schedule on every tick. Always shippable feasible Every returned plan was feasible and audited — a defensible answer on the wire inside the window, every tick. Why PRISM ### What changes when the solver fits the deadline. The same dispatch problem, three ways to solve it. The difference isn't a few percent of speed — it's whether you can solve the real fleet at all, in time. [CACHED · RTX 4000 Ada · NYISO Jan 2024] Standard QP solver Distributed methods PRISM (GPU) Largest fleet inside the 5-min window ~10,000 devices ~10,000 (slow) 1,000,000 1,000,000-device solve out of memory misses deadline ~15 s · 25× e2e Solution quality optimal — when it fits approximate ≤0.2% from optimal Feasibility guarantee yes not always certificate every solve Coordination across feeders small fleets only limited full · 12–28% uplift Sustains real-time, all day no marginal 288 / 288 cycles "Out of memory" is for a standard QP formulation on a single commodity card. PRISM's claim is the 10⁵–10⁶-device, 5-minute regime — not that no solver could ever do it. The assets under management Grid-scale batteries, home storage, EVs and solar — hundreds of thousands of them, each a decision PRISM makes every five minutes. Live · Run it yourself ### The dispatch engine, running in your browser. Not a video, not a mock-up — a full, executable research notebook with the governing equations, the constraints, real NYISO price data, and an interactive scenario explorer. No account, no sign-up, no install — it runs entirely in your browser, and nothing you type leaves your machine. Open fullscreen ↗ Open in Colab (Google sign-in) asymmetrycomputing.github.io/prism-vpp-demo · PRISM_VPP_Demo.ipynb Live kernel Booting Python in your browser… Pyodide runtime · first load ~10 s, then instant Python 3 (Pyodide, in-browser CPU) · use Run ▸ Run All Cells to recompute the economics with your own slider values · this notebook visualises validated results — the GPU solve times are cached measurements from a dedicated RTX 4000 Ada (the engine is never uploaded; a live GPU run is part of a pilot) · open in Colab to run it on free cloud CPU ✓ No account needed — runs in your browser Governing equations & constraints — disclosed Solver internals — black box Real NYISO price data · Jan 2024 1M-device solve in ~15 s Coordination premium 12–28% Every number labelled CACHED / SYNTHETIC / ASSUMPTION Interactive · Illustrative model ### Scale until the conventional solver gives up. Drag the fleet size and feeder constraint to see PRISM's measured solve time against the real 5-minute clearing window — and the coordination value unlocked once conventional QP solvers run out of memory. Built for the control room PRISM slots into the dispatch desk: a feasible, audited plan on the wire every tick — not a research prototype. 05 · What you get ### Built for an operations desk, not a demo. ✓ Deadline-bounded — the worst case is the deadline, not a solver timeout that misses the tick. ✓ Feasible by construction — a plan you can actually dispatch, every time, not one you have to repair. ✓ Stable — only the intervals worth changing move, keeping operations quiet. ✓ Auditable — content-hashed, reproducible records for the operator and the regulator. ✓ Deterministic — same inputs, same plan; re-derivable for any after-the-fact review. ✓ Integrates — a clean API that slots into your control loop and telemetry. Your fleet, from above Rooftop solar, EV-charging canopies, home batteries, a community storage compound — thousands of assets across one distribution network. Who it's for ### If you dispatch a fleet against a deadline, this is for you. VPP & aggregators #### More value per device Capture the 12–28% coordination premium across congested feeders that independent, device-by-device dispatch leaves on the table. Utilities & DSOs #### Feasible under real limits Dispatch that respects feeder, ramp and SoC constraints by construction — with a content-hashed audit trail for every plan. DERMS & OEM platforms #### Drop-in optimisation core Embed PRISM behind a clean API as the real-time engine inside your own platform. SDK and licensing available. FAQ ### Questions, answered plainly. Is the live demo really running, or is it a recording? It's a real, executable notebook running Python in your browser (via Pyodide) — no account, no install, nothing uploaded. The charts and tables you see are pre-computed so the page loads instantly; press Run ▸ Run All Cells to recompute them live with your own slider values. Does the demo run PRISM on a GPU? No — and we're explicit about that. The browser and Colab run on CPU and only visualise validated results. The headline solve times were measured offline on a dedicated RTX 4000 Ada GPU and are shown as cached benchmarks. The PRISM engine itself is never uploaded to a public cloud; a live GPU run on your data is part of a pilot. What exactly is the 12–28% coordination premium? It's the extra arbitrage value from optimising the whole fleet together — rather than each device on its own — when a distribution feeder is a binding bottleneck. It grows as the feeder tightens (12.75% → 19.03% → 28.09%) and is exactly 0% when no constraint binds, a control that validates the measurement. How is this different from the solver we already use? Standard QP solvers are excellent up to ~10k devices, then run out of memory; distributed methods scale further but miss the deadline. PRISM solves the full 10⁵–10⁶-device fleet inside the 5-minute window — a million devices allocated and disaggregated to feasible per-device setpoints in ~15 s, within ~0.2% of optimal — with a feasibility certificate on every solve. See the comparison table above. What do you disclose, and what stays proprietary? We disclose the entire problem — objective, decision variables and every constraint — because that mathematics is standard and public. We do not disclose how PRISM solves it. The engine is a black box with a clean input/output contract; full methodology is available under NDA. How do we try it on our own data? An 8-week paid pilot: you set the pass/fail metric before we start, we benchmark on your real fleet and market, and you get a report plus reference pricing for production. Every losing case is shown. ### Bring your tick. We'll hit it on your data. 8-week paid pilot on your data — you get a benchmark report on your real problem and reference pricing for production. You set the pass/fail metric before we start; every losing case is shown. Start an 8-week paid pilot → Prefer a form? Request a pilot → --- # Real-Time Portfolio Decisioning Source: https://asymmetrycomputing.com/applications/portfolio-decisioning Application · Real-Time Risk ## Rebalance and attribute risk inside the budget. Optimal rebalancing plus full risk attribution inside a hard latency budget — at universe sizes where standard optimizers stall. The decision isn't useful if it arrives after the moment to act has passed. PRISM returns an optimal rebalance, a factor/specific risk breakdown, and an audit trail — well under a second, across tens of thousands of holdings. In plain English. A large investment portfolio drifts out of shape as prices move, so it periodically needs a list of what to buy and sell to bring it back in line — while staying inside its risk limits. This page is about doing that across tens of thousands of holdings, fast enough to act on, and being able to show exactly where the portfolio's risk is coming from. Unfamiliar words? See the glossary . Start an 8-week paid pilot → All applications <1 s Optimal rebalance + risk attribution returned, well under a second 75 K Real-asset universe — tens of thousands of holdings factor +specific Full risk attribution returned alongside the trades audit A content-hashed trail returned with every decision PRISM · Real-Time Portfolio Decisioning A 75,000-asset rebalance with full factor and specific risk attribution , returned in well under a second — the trades and the reason for them, inside the desk's latency budget. ### The window is the budget. Scale is the problem. A rebalance you can't explain is a liability, and one that misses the latency budget is just history. The job is to return the optimal trades and the risk attribution together , inside the budget, at a universe size where standard optimizers stall — not to pick one of the three. Rebalance + risk attribution · output < 1s Factor / specific split shown is illustrative · full attribution returned per run 75,000 assets Returned in < 1 second 01 · The problem ### The right trade and the reason for it, before the moment passes. Real-time portfolio decisions carry two demands at once: produce the optimal rebalance , and produce the risk attribution that justifies it — both inside a latency budget set by the desk, not the solver. A great rebalance with no risk breakdown can't be signed off; a fast number that arrives after the window is closed can't be acted on. Now scale it. At tens of thousands of holdings, the universe itself becomes the obstacle: the work grows faster than the book, and standard optimizers stall exactly where the decision matters most. The constraint is doing all of it — optimal, attributed, auditable — at scale, in time. 02 · Why it's hard ### Optimal, attributed, and auditable — at scale, on a clock. Any one of these is routine. All of them together, at universe scale, inside a latency budget, is where conventional approaches run out of road. #### Scale is the obstacle At tens of thousands of holdings the universe grows faster than the book, and standard optimizers stall where it matters most. #### Trades and risk, together The rebalance and its factor/specific attribution have to come out of the same run, not a slow second pass that arrives too late to use. #### The budget is hard "Well under a second" is the spec, not the aspiration. A decision outside the latency budget is a decision you can't take. 03 · How PRISM fits ### A black box with a clean contract. You bring positions, targets, your risk view, and a latency budget; PRISM returns the trades, the attribution, and the audit trail — inside it. The methods are proprietary; the interface is simple. Inputs - Positions & targets - Your risk view - Cost assumptions - Hard latency budget → PRISM Real-time optimization core One engine, a hard budget, deterministic. → Outputs - Optimal rebalance - Factor / specific risk breakdown - Returned in well under a second - Content-hashed audit trail 04 · The evidence ### Demonstrated on a real 75,000-asset dataset. Demonstrated results on the dataset described — not a guarantee. Comparators are referred to generically as conventional / standard solvers. In the budget < 1s An optimal rebalance, a factor/specific risk breakdown, and an audit trail returned in well under a second. At scale 75,000 A real 75,000-asset universe — across tens of thousands of holdings, where standard optimizers stall. Explained & logged factor + specific The risk attribution comes out of the same run as the trades, with a content-hashed audit trail. Interactive · Illustrative model ### Price the cost of a late decision. Set your universe size and latency budget to see where a conventional solver misses the window — and the decision value PRISM protects by answering in time. 05 · What you get ### Built to be signed off, not just run. ✓ Sub-second at scale — an answer inside the latency budget at universe sizes where standard optimizers stall. ✓ Attribution included — a factor/specific risk breakdown out of the same run, not a slow second pass. ✓ Uses your risk view — keep the risk model you trust; upgrade the engine that acts on it. ✓ Auditable — a content-hashed, reproducible trail with every decision, for review and compliance. ✓ Deterministic — same inputs, same answer; re-derivable for any after-the-fact check. ✓ REST API + SDK — slots into your decisioning pipeline without a per-seat solver license. ### Your universe. Your budget. We'll return all three. 8-week paid pilot on your data — you get a benchmark report on your real problem and reference pricing for production. You set the pass/fail metric before we start; every losing case is shown. Start an 8-week paid pilot → Prefer a form? Request a pilot → --- # Direct Indexing & Tax-Aware Householding Source: https://asymmetrycomputing.com/applications/direct-indexing Application · Direct Indexing at Fleet Scale ## A fleet engine for personalized portfolios. Personalized, tax-aware portfolios across thousands of accounts — deterministic, auditable, and fast at fleet scale. Every account is its own problem; the promise is to re-solve all of them, correctly and on time, as a fleet. PRISM runs the batch and reports exactly how long it took, per percentile. In plain English. Instead of buying one fund that tracks an index, each client owns the actual shares in their own account. That allows the portfolio to be tailored to them and their tax bill to be actively managed — but it means every client needs their own calculation, every night. This page is about doing that for tens of thousands of accounts at once, before the market opens. Unfamiliar words? See the glossary . Start an 8-week paid pilot → Business view → 500 K Accounts in one book, priced in ~2 min on a single GPU ~25 × Faster than a tuned 10-core commercial/OSS deployment, matched quality p50·95·99 Full latency reporting on every batch — tax-lot handling built in audit A content-hashed trail for the whole fleet run PRISM · Direct Indexing at Fleet Scale Every account is its own tax-aware optimization. In a recent run, 500,000 accounts on a 1,000-name universe were priced in about two minutes on a single GPU — the whole book, before the open. ### Half a million accounts. One run. Reported to the percentile. At fleet scale, "it works" isn't enough — you need to know the p50, p95, and p99 of the run, because the tail is what blows the window. PRISM treats it as a production engine: batches from a thousand accounts up to a full 500,000-account book , each with tax-lot handling and a full audit trail. Fleet batch · per-account, at scale Batch Latency ladder is illustrative · p50 / p95 / p99 reported for every batch 1,000 10,000 500,000 01 · The problem ### Personalization is a promise to re-solve every account. A personalized, tax-aware account is a genuinely better product and a genuinely harder thing to run, because each account is its own optimization problem — its own lots, restrictions, and tax situation — that has to be re-solved as the world changes. Sell it to thousands of clients and you've signed up to run thousands of those problems, correctly, on a schedule. That makes it a fleet problem, not a single-portfolio one. The work grows with the number of accounts, not the dollars, and the thing that breaks the window isn't the average account — it's the tail. Which is why the run has to be measured the way production systems are: by percentile. 02 · Why it's hard ### Per-account correctness, fleet-scale throughput, tail you can trust. Each is manageable alone. Together — every account correct, the whole fleet on time, the tail bounded — is where most stacks hit a wall. #### Count is the multiplier Every account is a distinct problem and the count grows faster than assets — the work scales with the number of problems, not the dollars. #### The tail is the risk An average latency hides the accounts that blow the window. You need p50, p95, and p99 to trust the batch will finish on time. #### Taxes are per account Lot-level holding periods and wash-sale windows mean the right trade depends on each account's history, not just today's prices. 03 · How PRISM fits ### A black box with a clean contract. You bring the book; PRISM returns per-account trades, tax-lot handling, an audit trail, and the latency report for the batch. The methods are proprietary; the interface is simple. Inputs - Per-account positions & lots - Constraints & targets - Your risk view - Batch & deadline → PRISM Fleet optimization core One engine, fleet scale, deterministic. → Outputs - Optimized trades, per account - Lot-level tax handling - p50 / p95 / p99 latency report - Content-hashed audit trail 04 · The evidence ### A 500,000-account book, priced in about two minutes. Measured on a single RTX 4000 Ada GPU — 500,000 accounts across a 1,000-name universe, every configuration passing the tracking-error quality gate. Comparators are referred to generically as conventional / standard solvers. Whole book, cold start ~2 min 500,000 accounts optimized in ~134 s from cold on one GPU (~268 µs/account) — the full daily book, gate-passing quality. Intraday rescan ~1 min A warm re-solve of the same 500,000-account book in ~63 s (~126 µs/account) — fast enough to re-run the book intraday. vs a tuned baseline ~25–30× Against a tuned 10-core parallel commercial/open-source deployment (~56–68 min for the same book, extrapolated), at matched solution quality. How it's counted, honestly. The 500,000-account book was solved in chunks sized to the available card; a larger GPU means fewer chunks. PRISM's ~2 min is a fully measured wall-clock time; the baseline's ~56–68 min is extrapolated from a measured 2,000-account subsample of the same tuned parallel solvers — so we say "roughly 25–30×," not a spuriously precise figure. Every configuration passed the same tracking-error gate (max RMSE ≈0.001%, max TE ≈0.72 bps). The moat is the batched whole-book solve; a single isolated account is not where PRISM wins. Measured to the tail p50 · p95 · p99 Full latency reporting on every batch — the tail percentiles, not just an average that hides them. Correct & logged tax-lot Lot-level tax handling per account, with a content-hashed audit trail for the whole fleet run. Deterministic bit-exact Same inputs, same trades — re-derivable for any audit or exam date across the entire book. 500,000 accounts · one GPU · ~2 minutes The whole personalized book — exceeding a 400k-account programme — priced before the open on a single RTX 4000 Ada, at matched quality. Interactive · Illustrative model ### Size the tax alpha on your book. Move the sliders to your book, return, volatility, and tax rate to see the after-tax wealth a tax-loss-harvesting overlay can recover over time. 05 · What you get ### Built like production infrastructure. ✓ Fleet-scale — proven to a 500,000-account book on one GPU, with throughput that scales on the number of problems. ✓ Latency you can SLO — full p50 / p95 / p99 reporting per batch, so the tail is visible, not a surprise. ✓ Tax-aware — lot-level handling and wash-sale logic per account, built in. ✓ Deterministic — same inputs, same trades; re-derivable for any audit or exam date. ✓ Auditable — a content-hashed, reproducible record for the whole fleet run. ✓ Deploys your way — dedicated cloud, your VPC, or on-prem, with no per-seat solver license. ### Bring your book. We'll run the fleet and show you the tail. 8-week paid pilot on your data — you get a benchmark report on your real problem and reference pricing for production. Batched to your scale, with full latency reporting and every losing case shown. Start an 8-week paid pilot → Prefer a form? Request a pilot → --- # FRTB Standardised-Approach Capital Source: https://asymmetrycomputing.com/applications/frtb-capital Tool · FRTB SA Capital ## Price your FRTB capital, then optimize it. Under the standardised approach, market-risk capital is set by your sensitivities — and the charge is a hard, non-convex function to minimize against your hedges and limits. PRISM optimizes the true standardised charge within your risk limits, deterministically and with a full audit trail. Start with the interactive model below, then run it on your real book. In plain English. Banks are required by law to hold a cushion of their own money against the risks they take. The rules fix the formula for calculating how much — but not how a bank arranges its positions within that formula. This page is about that arrangement, which is worth real money because capital is expensive. Unfamiliar words? See the glossary . Start an 8-week paid pilot → All applications SA Standardised approach — the sensitivities-based method (SBM) true Optimizes the real non-convex charge, not a convex proxy limits Within your hedging and risk limits — never outside them audit Deterministic, content-hashed, re-derivable for the regulator PRISM · FRTB Standardised-Approach Capital Market-risk capital is a hard, non-convex function of your sensitivities. PRISM minimizes the true standardised charge within your hedging and risk limits — deterministically, with a regulator-ready audit trail. Interactive · Illustrative model ### Where your standardised charge comes from. Set your book notional and risk-class mix, then move hedge effectiveness and cross-class diversification to see how the standardised charge responds. Illustrative, not a regulatory calculation — your numbers will differ. How PRISM fits ### A black box with a clean contract. You bring your sensitivities, hedges, and limits; PRISM returns a capital-minimizing position within them, plus a deterministic audit trail. The methods are proprietary; the interface is simple. Inputs - Risk-class sensitivities - Available hedges - Risk & trading limits - Desk constraints → PRISM Capital optimization core The true SA charge, minimized within your limits. → Outputs - Capital-minimizing positions - Charge by risk class - Within every limit - Content-hashed audit trail Achievable capital relief depends entirely on your book, hedges, and limits — some books have little slack, others meaningful headroom. We don't quote a number we can't stand behind: the pilot measures it on your real positions, with losing cases shown. ### Bring your book. We'll show you the capital it's really carrying. 8-week paid pilot on your data — you get a benchmark report on your real book and reference pricing for production. You set the pass/fail metric before we start; every losing case is shown. Start an 8-week paid pilot → Prefer a form? Request a pilot → --- # Benchmarks & Evidence Source: https://asymmetrycomputing.com/benchmarks Benchmarks & Evidence ## Proof, not adjectives. Price the whole book before the open. PRISM by Asymmetry Computing is the tax-aware optimization core that scales to an entire personalized book — deterministically, auditably, with no per-seat solver license. Every figure below is measured on real US-equity data (March–April 2026), stated with its regime, with losing cases kept in. The peer-reviewable write-up is arXiv:2606.23367 , the evaluation artifacts are public at github.com/AsymmetryComputing , and the provenance of each number below is indexed on the research page . Run it on your data → How we benchmark Research & citations 10.2 × Fleet throughput vs a commercial CPU baseline (144 vs 14 accounts/sec/core) 11.6 min To price 100,000 personalized accounts on one core (vs ~118 min) 100,004 Real assets backtested at ~1.8s/rebalance · Sharpe 0.737 $238 k Full harvestable tax budget captured on a $5M, 192-name book Evidence integrity. Every number traces to one recorded benchmark run on real US-equity data. Comparators are labeled generically. We keep losing cases in — no cherry-picking, no stitched numbers, no parity we didn't measure. Measured, not asserted Every figure on this page traces to one recorded run on a dedicated GPU — real data, losing cases kept in. 00 · Verify it yourself ### The one benchmark you can run without us. Most numbers here are on real US-equity data we assembled. This one is on public Kenneth French (FF30) data anyone can download — so you can reproduce it on your own machine, with no licensed data and nothing to take on faith. Public-data benchmark 23.28× Faster than a commercial baseline at matched quality, on the public Fama–French 30-industry (FF30) dataset. Reproducible — no licensed data required. Nothing to take on faith public The inputs are the standard Fama–French dataset the whole field already knows — nothing bespoke, nothing cherry-picked. Then on your data 8-wk An 8-week paid pilot runs the same comparison on your real book, constraints, and deadline — every losing case shown. Latest · extreme scale · July 2026 ### The newest runs, at the frontier of scale. The most recent recorded benchmarks push past prior ceilings — a half-million-account book and a million-device fleet, each returned inside its operational window, at matched quality. One engine, finance to grid. Direct indexing · 500,000 accounts ~2 min A 500,000-account book on a 1,000-name universe priced in ~134 s cold / ~63 s warm on a single GPU — roughly 25–30× a tuned 10-core commercial/open-source deployment (extrapolated from a measured 2,000-account subsample), every configuration gate-passing. Real-time dispatch · 1,000,000 devices ~15 s A million-device fleet allocated and disaggregated to feasible per-device setpoints in ~15 s — 25× end-to-end vs a tuned commercial solver, within ~0.2% of optimal, well inside the 5-minute market window. Same engine, both 1 engine The same optimization core spans finance and grid — deterministic, feasible by construction, and auditable end-to-end at every scale on this page. Extreme-scale runs are chunked to the available card; PRISM times are fully measured wall-clock, baselines extrapolated where noted — so we say "roughly," not a spuriously precise figure. The batched whole-book/whole-fleet solve is the moat, not any single isolated problem. 01 · Scale incumbents can't reach ### The workflow exact solvers physically can't run. Nobody runs an exact solver across a 100,000-account book before the open — it times out long before. PRISM does, deterministically, and stays fast at every universe size. Full-market backtest 100,004 Real assets across 47+ exchanges, solved at ~1.8 seconds per rebalance. Walk-forward Sharpe 0.737, max drawdown −4.22%. Exact baseline ceiling ~5,000 Where an exact commercial solver times out (~33 s). PRISM prices the full book past that ceiling with no timeout. Fastest measured 190–40k PRISM is the fastest engine measured at every universe size across this range — not a single cherry-picked point. 02 · Fleet throughput ### ~10× more accounts per core — the margin lever. Personalized accounts grow faster than AUM. Throughput per core decides how many accounts an ops team can run, and whether the nightly batch clears on time. PRISM accounts / sec / core 144 Commercial CPU Baseline 1 accounts / sec / core 14 100,000 accounts · PRISM 11.6 min Priced on a single core — and it parallelizes linearly across cores. 100,000 accounts · baseline 118 min The same book on the commercial CPU baseline, single core. Determinism bit-exact Repeated runs return content-hashed, identical outputs — no solver timeouts, no flaky reruns. 03 · Tax alpha ### The only engine still harvesting six figures at scale. As beta commoditizes, after-tax outcome is how direct-indexing and SMA desks compete. This is where PRISM is measured strongest — and where simpler approaches quietly leave money on the table. Approach | Tax budget captured · $5M, 192 names | Six-figure alpha at 5,000+ names? | After-tax win rate vs PRISM PRISM | ~$238k (full budget) | Yes | — Open-source solver baseline | ~$3k at scale | No | PRISM wins ~88% Rules-based tax-loss harvesting | partial | No | PRISM wins ~95% Tax figures are measured on real-calibrated books; the robust real metric (tax captured) is measured directly. Your numbers are computed live on your data during the pilot. 04 · Quality, honestly ### Near-exact — and we name where we don't win. A sophisticated buyer trusts the vendor who states their non-fit first. So here it is, both sides. #### Where PRISM wins - Matches the exact reference optimum to ~1e-7. - Competitive with an exact commercial baseline on after-tax outcome: matches it in 58–63% of regimes, wins outright in ~33%. - Runs the fleet-scale and full-universe workflows where exact solvers can't run at all. - Deterministic, reproducible, auditable — every trade traces to your constraints and tax logic. #### Where PRISM does not - It does not beat the exact single-account optimum every time — no heuristic does. - It is not a bull-market beta play; diversification and tax discipline have a cost. - The quantum lane is a research track that adds zero tax alpha today — the production results are classical and measured. 05 · The measured curves ### Every headline number, charted. The same evidence behind the claims above, plotted straight from the recorded runs — factor-constrained QP on real US-equity universes, March–April 2026, p50 of five timed runs. Lower is better on solve time; speedup is versus the commercial baseline on CPU. Figure 1 · Solve time vs universe size — factor-constrained QP PRISM (GPU) Commercial baseline (CPU) Open-source solver 1 (CPU) PRISM stays lowest across the universe sizes tested, at sub-ppm quality gaps against the reference. Solver Runtime p50 (ms) Gap vs reference Stability (CV) PRISM (GPU) | 138.7 | < 0.00001% | 11.7% Commercial baseline (CPU) | 208.1 | reference | 19.2% Open-source solver 1 (CPU) | 466.9 | < 0.00001% | 10.0% Open-source solver 2 (CPU) | 768.6 | < 0.00001% | 12.0% GPU baseline | 870.0 | 0.0014% | 8.0% Open-source solver 3 (CPU) | 11,311 | < 0.00001% | 10.0% Figure 3 · Structured QP at 5,000 real assets. PRISM is the fastest exact-quality solver in the field — 1.5× under the commercial baseline and 3.4× under Open-source solver 1 — while staying within a tight stability band. We keep the honest qualifier in: GPU baseline posts a lower run-to-run CV, and the commercial baseline remains the correctness reference every other engine is measured against. Figure 4 · Fleet throughput — personalized accounts priced per second, per core PRISM (GPU) ~11.6 min for 100,000 accounts 144 acct/s Commercial CPU baseline ~118 min for 100,000 accounts 14 acct/s ~10.2× per-core throughput, and PRISM parallelizes linearly across cores. The whole personalized book is priced inside the overnight window, before the open — the workflow the fleet is actually billed on. Solver | QP Speed | Replay | Transition | Quality | Stability | Max Scale PRISM (GPU) | 10 | 10 | 10 | 10 | 9 | 10 Commercial baseline (CPU) | 7 | 6 | 3 | 10 | 7 | 8 Open-source solver 1 (CPU) | 6 | 5 | 0 | 9 | 7 | 5 GPU baseline | 5 | 2 | 1 | 2 | 6 | 5 Open-source solver 3 (CPU) | 2 | 2 | 1 | 3 | 5 | 4 Open-source solver 2 (CPU) | 4 | 1 | 0 | 9 | 3 | 4 Figure 6 · Competitive overview — six solvers across six dimensions, scored 0–10 on real-data benchmarks (March 2026); higher is better. PRISM leads on speed, transition workflows, and scale; the commercial baseline and Open-source solver 2 match it on single-problem quality; GPU baseline edges run-to-run stability. No solver wins every axis — which is the point of showing all six. 06 · Trust primitives ### Built to pass compliance, not just benchmarks. In this buyer set, trust is the product and compliance is the gate. PRISM ships the de-riskers up front. #### Deterministic & reproducible Content-hashed, bit-reproducible outputs — re-derivable for any audit or exam date. #### Tax-rule correctness Wash-sale handling and lot-level tax accounting, validated by an internal test suite — 18 tests pass. #### Validate against ground truth An exact-reference comparator ships so you can check PRISM against the exact optimum on your own data. #### Transparent, not a black box Every trade traces to explicit constraints and tax logic — no opaque ML in the trade path. 07 · How we benchmark ### Matched workloads. Honest qualifiers. Losses kept in. A benchmark is useful only when it tells you whether an engine can clear your workload with acceptable quality, latency, and reliability — not when it collapses to one headline number. Real data Core results run on real US-equity data and a real 100,004-asset benchmark — full production scale. Matched comparison Same universe, constraints, costs, and tax rules across every engine. Quality gap reported alongside runtime. Reproducible Fixed seeds, recorded runs, low run-to-run variance. Every figure ties to an artifact. Priced in decisions Cost per solve at published cloud rates runs 4–27× below tuned CPU incumbents — a full 500,000-account book is about 3¢ of GPU time. The economics → On comparators: baselines are labeled generically — Commercial CPU Baseline 1 / 2 (exact and fast-lane commercial solvers), open-source solver baselines , and rules-based tax-loss harvesting . We compare on outcomes (after-tax wealth, throughput, scale, feasibility), state the regime, and include the cases where a baseline wins. PRISM's results describe outcomes; its methods are proprietary. Read more: how to read optimization benchmarks without being misled → ### The proof that counts is on your data. A 30-day, buyer-owned matched-workload pilot: PRISM vs your current stack, on your universe, constraints, costs, and tax rules — a pass/fail metric you set before we start. You get a full results pack (every account, losses shown), deterministic audit logs, and an ROI computed with your real numbers. Request a matched-workload pilot → --- # Glossary — Plain-English Definitions Source: https://asymmetrycomputing.com/glossary Glossary ## Every term on this site, in plain English. Asymmetry Computing builds PRISM, an optimization engine for institutional finance and energy. Those fields have a lot of vocabulary, and most of it is simpler than it sounds. This page defines every specialist term used across the site — no prior background assumed. If something here is still unclear, that is our failure, not yours: tell us at debdoot@asymmetrycomputing.com . Start here ### The one-paragraph version. Some decisions have three properties at once: there are far too many possible answers to check by hand, the answer has to obey strict rules, and it has to arrive before a specific moment or it is useless. Deciding what to buy and sell across thousands of investment accounts before the market opens is one. Deciding what every battery on an electricity grid should do in the next few seconds is another. PRISM is software that answers questions like these inside the time available, and shows its working. The rest of this page explains the words the industry uses for the pieces. The core idea ### What optimization actually is, in ordinary language. #### Optimization Picking the best option out of an enormous number of possibilities, when the options have to obey rules. Not "making something a bit better" — literally searching a space too large to check by hand and returning the best choice available. Why it matters here Everything on this site is a version of this one problem. #### Constraint A rule the answer is not allowed to break. "Hold no more than 5% in any one company." "Do not sell anything bought in the last 30 days." "Never discharge the battery below 20%." Why it matters here Constraints are what make the problem hard. Without them, most of these questions would be easy. #### Feasible An answer that breaks none of the rules. A feasible answer might not be the best one, but it is legal and usable. Why it matters here An infeasible answer is worse than a slow one — it can't be acted on, so a person has to step in and fix it by hand. #### Near-optimal Extremely close to the mathematically perfect answer, but not provably identical to it. In practice the difference is far smaller than the noise in the inputs. Why it matters here For a decision that has to be acted on, a difference smaller than the uncertainty in the inputs is not a real difference. #### Solver The software that does the searching. You hand it the problem and the rules; it hands back an answer. Why it matters here PRISM is a solver built for problems that must be answered by a fixed moment in time. #### Deadline-bounded A problem where the answer is worthless if it arrives late, even if it is perfect. The market opens; the grid tick closes; the trading window shuts. Why it matters here This is the whole idea behind PRISM. The deadline is a hard rule, not a preference. #### Deterministic Same inputs, same answer, every single time. Run it again next year and you get exactly what you got today. Why it matters here Regulators and auditors need to reproduce a decision. "It came out slightly different" is a finding. #### Audit trail A record of why a decision was made — which rules applied, what the inputs were, what came out. Why it matters here In regulated finance, "the computer decided" is not an acceptable explanation. Portfolios and tax ### Terms from wealth and asset management. #### Direct indexing Instead of buying one fund that tracks an index, you buy the individual shares that make up that index, in your own account. It costs more to run, but because you own the actual shares you can customise and manage tax on each one. Why it matters here It creates millions of individual portfolios that each need their own calculation — the workload PRISM was built for. #### Householding Treating all the accounts belonging to one family as a single picture — joint accounts, IRAs, trusts, each child's account — rather than optimising each one in isolation. Why it matters here Much harder, because a good decision in one account can be a bad one for the household. #### Tax-loss harvesting Deliberately selling an investment that has lost value to book the loss, which offsets tax owed on gains elsewhere, then buying something similar to stay invested. Why it matters here Real money. Doing it well across thousands of accounts is a large, repetitive calculation. #### Harvestable tax budget The maximum tax benefit actually available to capture in a given account over a period. It is a ceiling, not a target. Why it matters here "Captured the full $238K budget" means all of what was available was collected — not that $238K was created out of nowhere. #### Wash sale A tax rule: if you sell at a loss and buy the same or a very similar investment within 30 days, you don't get to claim the loss. Why it matters here It turns a simple sell decision into a rule with sharp edges and a memory. #### Tax lot Each separate purchase of the same investment, tracked individually with its own purchase date and price. 100 shares bought in three batches is three lots. Why it matters here Which lot you sell changes the tax bill, so the choice is part of the optimization. #### Rebalancing Adjusting a portfolio back toward its intended mix after prices drift it out of shape. Why it matters here Routine for one account. At tens of thousands of accounts, against a deadline, it is an infrastructure problem. #### Book All the accounts or positions a firm manages, taken together. "Pricing the whole book" means computing every account. Why it matters here The book, not the single account, is the real unit of work. #### Universe The full set of investments the optimizer is allowed to choose from. A 75,000-asset universe means 75,000 candidates. Why it matters here The larger the universe, the harder the problem grows. #### Turnover How much of a portfolio is bought and sold in a period. High turnover means more trading, more cost, more tax. Why it matters here Usually capped, which is another constraint. #### Tracking error How far a portfolio drifts from the index it is meant to follow. Low tracking error means it behaves like the index. Why it matters here The tension at the heart of direct indexing: customise and manage tax, without drifting too far. #### Risk attribution Breaking down where a portfolio's risk actually comes from — how much from the whole market, an industry, a style, and how much specific to individual holdings. Why it matters here Tells you what you're actually exposed to, rather than just the total. #### Factor A shared force that moves many investments together — the overall market, company size, industry, value versus growth. Why it matters here Factors are why a portfolio can look diversified and still be exposed to one thing. #### SMA Separately Managed Account. A portfolio owned directly by one client and managed to their situation, rather than a pooled fund shared with strangers. Why it matters here The vehicle direct indexing is usually delivered through. #### After-tax alpha The extra return that survives after tax. A strategy that beats the market before tax and loses to it after tax has no after-tax alpha. Why it matters here The number a taxable investor actually keeps. #### Basis point One hundredth of one percent. 50 basis points = 0.50%. Why it matters here Standard shorthand in finance for small differences that compound into large sums. Speed and reliability ### How response time is measured honestly. #### Latency How long something takes to respond, from asking to getting an answer. Why it matters here For a deadline-bounded workload, latency is the product. #### p50, p95, p99 Percentiles. p50 is the middle: half of runs were faster. p95 means 95% were faster and 5% were slower. p99 is the slowest 1%. Why it matters here Averages hide the failures. A system with a fine average and a terrible p99 misses deadlines where it hurts. #### Tail latency The slow outliers — the p95 and p99 cases. Why it matters here A batch finishes when its slowest member finishes, so the tail sets the real completion time, not the average. Energy and the grid ### Terms from the electricity side. #### Dispatch Deciding what each power resource should do right now — charge, discharge, hold, produce more, produce less. Why it matters here Must be recomputed continuously as prices and demand move. #### DER Distributed Energy Resources. Small power assets spread across the grid instead of one large plant: rooftop solar, home and commercial batteries, EV chargers, backup generators. Why it matters here Thousands of small assets to coordinate rather than a handful of big ones. #### VPP Virtual Power Plant. Many distributed resources coordinated so that together they behave like a single power plant. Why it matters here Structurally very similar to managing a portfolio of accounts — which is why one engine serves both. #### State of charge How full a battery is right now. Why it matters here It links time periods together: energy used now is unavailable later, so the intervals can't be solved separately. #### Control tick The fixed heartbeat of a grid control system. Each tick it looks at the current state and issues new instructions. Why it matters here The tick does not wait. Miss it and the previous plan stands. #### CAISO / ISO An ISO (Independent System Operator) runs the electricity grid for a region and matches supply to demand. CAISO is California's. Why it matters here Real CAISO data is what the energy dispatch results were measured on. Regulation ### Terms from bank capital rules. #### Regulatory capital The cushion of its own money a bank is required by law to hold against the risks it takes, so that losses hit the bank's owners before its depositors. Why it matters here Capital is expensive, so how it is calculated matters enormously. #### FRTB Fundamental Review of the Trading Book. The international rules governing how banks calculate the capital they must hold against trading risk. Why it matters here It changed the calculation, which changed what banks have to hold. #### Standardised approach The regulator's prescribed formula for calculating required capital, as opposed to a bank's own internal model. Why it matters here The formula is fixed, but how a portfolio is arranged within it is not — and that is an optimization problem. ### Still not sure this applies to you? The quickest way to find out is to describe your workload in your own words. If it isn't a fit, we will say so — we publish where we don't win too. Talk to us → --- # Research & Citations Source: https://asymmetrycomputing.com/research Research & citations ## The published record behind every number we quote. Asymmetry Computing publishes the evidence for PRISM in three places: a peer-reviewable paper on arXiv, public evaluation artifacts on GitHub, and the benchmark pages on this site. This page is the index to all of it, with permanent links. If you want to check a claim we make, start here. Read the paper on arXiv → Evaluation artifacts on GitHub 01 · Peer-reviewable research ### The paper. Asymmetry PRISM: A CPU/GPU Portfolio Optimization Engine for Deadline-Bounded Institutional Rebalancing. Debdoot Ghosh, Asymmetry Computing. June 2026. arXiv:2606.23367 . The paper treats institutional rebalancing as what it actually is operationally: a batched optimization workload with a hard operating deadline, where hundreds of accounts need new weights under budget, turnover, exposure, exclusion, and tax-aware controls before trading can proceed. PRISM is evaluated across a public evaluation boundary — problem data in; returned weights, status codes, timings, memory class, external feasibility diagnostics, eligible objective comparisons, and audit records out. Three results are reported, in the paper's own terms: - On completed multi-solver rows from N=100 to N=2,000, Asymmetry PRISM-CPU is 4.5× to 24.1× faster than the fastest completed reference row in the same lane. - In the production queue study, Asymmetry PRISM-GPU completes 500/500 accounts over a 10,000-instrument universe in 109.5 s within a declared 25-minute operating window, with zero missed deadlines and an audit record for every solve. The recorded open-source solver queue baseline completes 4/500 . - On an operationally constrained real-data suite (tax-motivated transition penalties, restriction caps, turnover controls, batches), Asymmetry PRISM clears constrained solves 3.4× to 126.7× faster than the best completing incumbent at certified-equal objectives, with the GPU route widening to 8.8× over the CPU route at N=384,800. Rows without a completed reference are reported as feasibility, timing, memory, and failure-status evidence rather than being dropped or scored as an infinite speedup. Abstract · PDF · Full text (HTML) 02 · Public code ### The evaluation artifacts. Result tables, evidence ledger, and external feasibility and residual checks — published so a buyer can audit the claims without us in the room. #### prism-public-evaluation Public evaluation artifacts for the Asymmetry PRISM benchmark: result tables, the evidence ledger, and external feasibility and residual checks that reproduce the figures reported in arXiv:2606.23367. #### prism-di-public-evaluation The direct-indexing evaluation repository: batched, tax-aware account solves with per-percentile latency reporting, published as a runnable evaluation harness. #### github.com/AsymmetryComputing The organization index — every public repository Asymmetry Computing maintains, including the interactive demo notebooks linked from the applications pages. 03 · Claim provenance ### Every headline number, and where it comes from. Comparators are described generically. Every figure traces to a recorded run on real data, and the losing cases stay in. See the full evidence → Claim | What it measures | Where to verify ~5 ms CAISO dispatch A feasible, audited grid dispatch plan returned within deadlines from 500 ms down to ~5 ms, on real California ISO data. Energy dispatch 75,257-asset rebalance in under a second A full rebalance with factor and specific risk attribution plus an audit trail, at a universe size where standard optimizers stall. Portfolio decisioning $238K tax budget captured The full harvestable tax budget captured on a $5M direct-indexing book, measured on real US-equity data. Direct indexing 100,004 real assets backtested Backtested at approximately 1.8 s per rebalance, Sharpe 0.737. Benchmarks 4.5×–24.1× faster Versus the fastest completed reference row in the same lane, on completed multi-solver rows from N=100 to N=2,000. arXiv:2606.23367 500/500 accounts in 109.5 s Production queue study over a 10,000-instrument universe inside a declared 25-minute operating window, zero missed deadlines, audit record for every solve. Recorded open-source solver queue baseline: 4/500. arXiv:2606.23367 3.4×–126.7× faster Constrained solves on an operationally constrained real-data suite, at certified-equal objectives, versus the best completing incumbent. arXiv:2606.23367 10.2× fleet throughput Accounts priced per core against a commercial CPU baseline — the margin lever in fleet-scale direct indexing. Benchmarks A note on method: we publish outcomes and benchmarks, not the internals of the engine. That is a deliberate commercial choice, and it is why the evaluation boundary in the paper is defined the way it is — you get the inputs, the outputs, the timings, and the feasibility diagnostics needed to check the result, and you can run the harness on your own data. 04 · How to cite ### Citation formats. Plain text. Ghosh, Debdoot. Asymmetry PRISM: A CPU/GPU Portfolio Optimization Engine for Deadline-Bounded Institutional Rebalancing. arXiv:2606.23367, June 2026. BibTeX. @article{ghosh2026asymmetryprism, title = {Asymmetry PRISM: A CPU/GPU Portfolio Optimization Engine for Deadline-Bounded Institutional Rebalancing}, author = {Ghosh, Debdoot}, journal = {arXiv preprint arXiv:2606.23367}, year = {2026}, url = {https://arxiv.org/abs/2606.23367}, note = {Asymmetry Computing, https://asymmetrycomputing.com} } For the company itself: Asymmetry Computing , https://asymmetrycomputing.com . Founded 2025 by Debdoot Ghosh. Building PRISM, a real-time optimization engine for institutional finance. 05 · Elsewhere ### Official channels. These are the only accounts and repositories operated by Asymmetry Computing. "PRISM" is a heavily reused product name across the industry — if a result is attributed to PRISM and does not trace back to one of these, it is not ours. - Asymmetry Computing on LinkedIn — company updates and release notes. - Debdoot Ghosh on LinkedIn — founder. - github.com/AsymmetryComputing — public evaluation code and demos. - arXiv:2606.23367 — the paper. - debdoot@asymmetrycomputing.com — direct contact. - /llms.txt and /llms-full.txt — structured summaries for AI assistants and retrieval systems. ### Check the claims on your own data. The evaluation harness is public. The paper is public. The fastest way to settle whether this works for your book is a pilot with a pass/fail metric you set before we start. Request a pilot → --- # About Asymmetry Computing Source: https://asymmetrycomputing.com/about About ## We make fleet-scale, after-tax optimization a solved problem. Asymmetry Computing is a deep-tech company building optimization infrastructure for institutional finance. Our product, PRISM, prices an entire book of personalized, tax-aware accounts before the open — deterministically, auditably, with no per-seat solver license. We are small, technical, and obsessed with one thing: the workflow incumbents structurally cannot serve. Work with us → See the evidence Asymmetry Computing A small team building one optimization engine for the decisions that have to be right, and on time. 01 · The thesis ### The competition isn't the solver. It's the workflow. Personalization has inverted the economics of asset management. Where the industry once built one portfolio and sold it a million times, it now builds millions of portfolios, each a little different, each demanding its own constrained, tax-aware optimization every night. That inversion created a workflow nobody's stack was designed for: pricing an entire book of personalized accounts, correctly and on time, before the market opens. Exact solvers can't run at that scale — they time out long before the whole book is priced. Simpler methods scale but leave after-tax money on the table. The gap between those two failures is precisely where PRISM lives: near-exact, after-tax-optimal trades at fleet scale, with the determinism and audit trail that compliance requires and no license tax that punishes growth. We are not building a better solver. We are building the optimization core the whole industry is missing once account counts and universes grow. 02 · What we believe ### Honesty is a feature, not a constraint. The way we work is the reason sophisticated buyers trust us. #### Outcomes, not claims Every public number traces to a recorded run on real data. We publish results and benchmarks — the methods stay ours. #### We keep the losses in We show where we don't win — single-account exact optima, bull-market beta, quantum tax miracles. Naming our non-fit is how we earn the wins. #### Trust is the product Deterministic, content-hashed, re-derivable outputs. In a market where trust is the scarce resource, auditability isn't a checkbox — it's the point. 03 · The proof ### Measured, on real US-equity data. March–April 2026. Comparators labeled generically; losing cases kept in. See the full evidence → 10.2 × Fleet throughput vs a commercial CPU baseline 100,004 Real assets backtested at ~1.8s/rebalance · Sharpe 0.737 $238 K Full harvestable tax budget captured on a $5M book 18 Internal correctness tests passing; exact-reference comparator ships 04 · The company ### Founder-led, technical, and early — on purpose. Asymmetry Computing was founded in 2025 by Debdoot Ghosh and builds PRISM, a real-time optimization engine for institutional finance. The company is early-stage and founder-led, and it is transparent about that, because it shapes how we sell: not on a wall of logos we don't yet have, but on a buyer-owned pilot that runs on your data, with a pass/fail metric you set before we start. The first reference earns the trust; we are asking to be the engine that proves it on your book. The technical work is published rather than asserted. Debdoot is the author of Asymmetry PRISM: A CPU/GPU Portfolio Optimization Engine for Deadline-Bounded Institutional Rebalancing (arXiv:2606.23367, June 2026), which frames institutional rebalancing as a batched optimization workload under a hard operating deadline and evaluates PRISM across a public evaluation boundary. The accompanying artifacts — result tables, the evidence ledger, and external feasibility and residual checks — are published at github.com/AsymmetryComputing , so a prospective buyer can audit the claims without taking our word for any of them. That is the whole posture of the company in one line: the result sells, the method stays ours, and everything we claim is checkable. A full index of the papers, repositories, and the provenance of each headline number lives on the research page , and the measured runs behind them are on benchmarks . If you run thousands of personalized, tax-aware accounts — or you build the infrastructure that does — and your current stack is hitting a wall on scale, determinism, or license cost, we should talk. Request a pilot → debdoot@asymmetrycomputing.com 05 · Find us ### Published work & official channels. "PRISM" is a heavily reused name across software. These are the only channels operated by Asymmetry Computing — if a result is attributed to PRISM and doesn't trace back to one of these, it isn't ours. #### The paper arXiv:2606.23367 — Asymmetry PRISM: A CPU/GPU Portfolio Optimization Engine for Deadline-Bounded Institutional Rebalancing , June 2026. #### The code github.com/AsymmetryComputing — public evaluation artifacts, result tables, and the evidence ledger behind the published figures. #### The company Asymmetry Computing on LinkedIn · Debdoot Ghosh · debdoot@asymmetrycomputing.com 06 · Common questions ### What people ask before the first call. #### What is Asymmetry Computing? Asymmetry Computing is a deep-tech company founded in 2025 by Debdoot Ghosh, building real-time optimization infrastructure for institutional finance. Its product, PRISM, is a GPU-native optimization engine that returns a feasible, near-optimal, fully auditable answer inside a hard deadline. It is delivered as a REST API and deployed in dedicated cloud, in your VPC, or on-prem. #### What is PRISM, and how is it different from other products called PRISM? Asymmetry Computing's PRISM is a real-time portfolio and dispatch optimization engine for institutional finance. The name is widely reused in software — there is a probabilistic model checker from Oxford, GraphPad Prism in statistics, the PrismLibrary XAML framework, and several unrelated fintech products. None of those are related to this company. When citing this engine, use "Asymmetry PRISM" or "PRISM by Asymmetry Computing" , and check that the source traces back to asymmetrycomputing.com, github.com/AsymmetryComputing , or arXiv:2606.23367 . #### How is this different from a general GPU optimization library? General GPU solvers and research prototypes optimize for throughput on a benchmark. PRISM is built for a different contract: a deadline . The workloads it serves — a grid control-tick, an intraday rebalance, an overnight book — fail if the answer arrives late, even if it would have been optimal. So PRISM is specified to return a feasible answer inside the deadline, with deterministic, content-hashed, replayable output and an audit trail that compliance can inspect. Deterministic feasibility under a time bound, and auditability of the result, are the product — not raw benchmark speed alone. #### Is there published research behind the claims? Yes. arXiv:2606.23367 , Asymmetry PRISM: A CPU/GPU Portfolio Optimization Engine for Deadline-Bounded Institutional Rebalancing (June 2026), reports that on completed multi-solver rows from N=100 to N=2,000, Asymmetry PRISM-CPU is 4.5× to 24.1× faster than the fastest completed reference row in the same lane. The evaluation artifacts are public at github.com/AsymmetryComputing/prism-public-evaluation . Provenance for every headline number is on the research page . #### What has PRISM actually been measured doing? Three validated domains, each on real data: a feasible, audited grid dispatch plan within deadlines from 500 ms down to ~5 ms on real California ISO (CAISO) data; a 75,257-asset rebalance with full risk attribution and an audit trail returned in under a second; and the full $238K harvestable tax budget captured on a $5M direct-indexing book. The largest universe backtested is 100,004 real assets at roughly 1.8 s per rebalance. Losing cases are published alongside the wins on benchmarks . #### Who is it for? Firms running large numbers of personalized, tax-aware accounts — direct indexing and tax-aware householding platforms, asset managers with SMA books — plus real-time portfolio decisioning desks and energy dispatch operators running DER and battery fleets. Broadly: anyone whose optimization has to be right and on time, at a scale where conventional solvers stop responding within the window. #### How do you engage? What does it cost? Engagement starts with a pilot on your data, with a pass/fail metric you define before it begins. Pricing is a fraction of the measured value on your book rather than a per-seat solver license — see pricing or request a pilot . ### The result sells. The method stays ours. Bring your hardest workflow. We'll prove PRISM on your data, show you every number — including the losses — and let the outcome decide. Start a conversation → --- # Pricing Source: https://asymmetrycomputing.com/pricing Pricing ## Priced to the value, not the compute. PRISM's worth is the after-tax alpha it preserves, the accounts it lets you run per core, and the solver license it eliminates — not CPU seconds. So that's how we price it. Land with a paid pilot credited toward production, then scale with a model that grows as your book does. No per-seat solver license, ever. Request pricing & a pilot → See the ROI evidence Priced on proof An 8-week paid pilot on your data decides it — a benchmark report and reference pricing for production, every losing case shown. Pilot #### Prove it on your data Flat paid fee · credited toward year-one production - 30-day, buyer-owned matched-workload pilot - Your universe, constraints, costs, and tax rules - Pass/fail metric agreed before kickoff - Full results pack — every account, losses shown - Deterministic audit logs + exact-reference comparator - ROI computed with your real numbers Start a pilot Production · most popular #### Scale with your book Account-volume or AUM-linked · declining unit price as you grow - Price the whole personalized book before the open - Account-volume tiers or low-bps-on-AUM, your choice - REST API + Python SDK into your nightly batch - Dedicated cloud or your VPC - Deterministic, reproducible, audit-ready outputs - No per-seat solver license — scale without a license tax Talk to us Enterprise #### Own it on your infra Annual enterprise license · uncapped accounts - On-prem or air-gapped deployment - Uncapped accounts, predictable annual cost - Data residency — your data never leaves - Model-governance pack + SOC 2 roadmap - Priority support and integration - OEM / embedded options for platforms Contact sales Reference scale: pilots are a fixed five-figure fee, credited toward production; production licenses are typically five-to-six figures per year and scale with your book. Exact pricing is scoped to your workflow during the pilot. The ROI frame ### The price should be a small fraction of the value. Five measured levers PRISM moves — we compute the exact multiple on your numbers during the pilot. Tax alpha preserved A few bps of incremental after-tax return across a tax-managed book is worth multiples of any tier. PRISM captures six-figure tax alpha where simpler methods capture thousands. Accounts per core ~10× throughput means running the book on far fewer cores with fewer manual interventions — more accounts per ops head. License & timeout cost No commercial solver license, and no missed-rebalance risk from timeouts in the night batch. We don't ship a fabricated ROI number. We ship the model and the measured anchors, then compute your specific return — AUM, account count, ops headcount, license spend, incident log — live during the pilot. A buyer-owned ROI number is worth more than any slide. ### Let's price it against the value on your book. Request a pilot and pricing. We'll scope it to your workflow, run it on your data, and show you the return — including every losing case. Request pricing & a pilot → --- # Trust & Compliance Source: https://asymmetrycomputing.com/security Trust & Compliance ## In this market, trust is the product. Compliance is the gate every institutional deal passes through, and most optimizers worry compliance teams because they can't reproduce or explain their own decisions. PRISM is built the opposite way: deterministic, re-derivable, and auditable by design. We bring the compliance conversation to the first call, not the last. Talk to our team → See the evidence bit-exact Deterministic, content-hashed outputs — re-derivable for any exam date 18 Internal correctness tests passing (incl. tax-rule tests) exact-ref Validation comparator ships — check us against ground truth on your data on-prem Deploy in your VPC or air-gapped — your data never leaves Deploy in your VPC, on-prem, or air-gapped Deterministic, content-hashed, and yours to run behind your own perimeter. 01 · Already built ### The trust primitives, shipping today. Not a roadmap promise — these are properties of the engine right now. #### Deterministic & reproducible Content-hashed, bit-reproducible outputs tied to a data hash. Any past decision can be re-derived exactly — the reproducibility problem was found and fixed, and every result ties to its inputs. #### Auditable & transparent Every trade traces to your explicit constraints and tax logic. It is a transparent optimizer, not a neural black box — there is no opaque ML in the trade path to explain away. #### Tax-rule correctness Wash-sale handling (IRC §1091, including the 61-day window and household scope) and specific-lot selection, validated by an internal test suite — tax-rule tests included in the 18 passing tests. #### Validate against ground truth An exact-reference comparator ships with the engine, so your own quants can confirm PRISM's quality against the exact optimum on your own data — trust by verification, not by faith. 02 · The narrative ### A compliance upgrade, not a risk. Most optimizers make compliance nervous for a simple reason: they are non-reproducible. If you cannot re-derive why a trade happened, you cannot defend it to an examiner, a client, or your own risk committee. PRISM inverts that. Because outputs are deterministic and content-hashed, any past decision is exactly reproducible; every trade traces to your constraints and explicit tax logic; the wash-sale handling is testable against your own policy; and you can validate the engine against an exact solver on your own data before you ever rely on it. The practical effect is that the team usually positioned as the blocker becomes an ally. Reproducible, re-derivable decisions cut audit and exam-prep time and shorten model-validation cycles. That is why we bring an audit-log sample, the validation methodology, and a model-governance outline to the first compliance conversation — pre-empting the objections is how the wedge deal actually closes. 03 · Roadmap ### Built to unlock the largest, most-regulated buyers. What's shipping now, and what we're formalizing as we scale into platforms, custodians, and large asset managers. Live #### Deterministic audit logs Re-derivable, content-hashed records for every run. Live #### On-prem / VPC / air-gapped Deploy where your data is allowed to live — no per-seat license. Roadmap #### SOC 2 (Type I → II) The procurement gate at platforms, custodians, and large AMs. Roadmap #### Model-governance docs SR 11-7-style model-risk documentation, assembled from the validation work and comparator. Roadmap #### Immutable audit-log export Per-decision export for exams and client reporting. Roadmap #### Kill-switch & guardrails Explicit hard-stop and constraint-violation tripwires for ops and compliance. 04 · The Trust Pack ### What we bring to your first compliance call. ✓ Deterministic-output proof and a sample audit log . ✓ Tax-rule (wash-sale & lot-selection) test results . ✓ The exact-reference comparator methodology for validating quality. ✓ The SOC 2 roadmap and model-governance outline. ✓ Data-handling and deletion terms for the pilot and production. ✓ The kill-switch / guardrail design for ops sign-off. ### Bring compliance to the first meeting. We'll be ready. Start a buyer-owned pilot and put the trust primitives to the test on your data — deterministic logs, validation against ground truth, and tax-rule correctness on your own policy. Request a pilot →