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.

500K
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
A vast grid of personalized investment accounts rendered as points of light
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.

A vast data-centre hall computing an entire book of accounts at once
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 →