API · Live now

The PRISM Optimization API.

Firm-wide portfolio optimization at book scale, as a hosted API. Bring accounts, lots, and constraints; get back optimized weights, timings, and a hash-chained audit trail — no solver license, no on-prem cluster, no queueing overnight batch.

Get a free trial key → Read the API docs
47s
500,000 accounts × 1,000 names rebalanced firm-wide — 94µs/account
14s
Firm-wide intraday what-if rescan of the same book
0.00025%
Max weight error vs. an independent open-source reference at 1e-9 tolerance, validated in-job (gate: 0.05%)
65ms/acct
The reference solver's own per-account time — roughly 9 hours single-threaded for the same book

Measured 2026-07-11 on a datacenter GPU. Independent validation runs inside every job — not a separate marketing benchmark.

01 · Built for real books

Every constraint your desk runs. One job.

PRISM handles the full complexity of a real personalized book — lots, taxes, households, wash sales, caps — in a single firm-wide pass.

✓Multi-lot, tax-aware direct indexing. HIFO/specific-lot basis tracking across thousands of accounts at once.
✓Householding scans. Optimize related accounts together without violating account-level constraints.
✓Tax-loss-harvest scans. Identify and execute harvestable losses across the book in the same pass.
✓Wash-sale constraints. IRS-compliant lookback enforced natively, not bolted on after the fact.
✓Embedded-gain handling. Legacy, concentrated positions are respected, not liquidated by default.
✓Taxable / tax-deferred mixes. Asset location handled correctly across account types in one job.
✓Per-account position caps. Concentration limits enforced per account, per name, at book scale.
✓Hash-chained audit trail. Every job returns a SHA-256 hash-chained record of every processing stage — built for compliance review, not an afterthought.
Compliance-friendly by default. Every job — not just the ones you remember to log — returns a deterministic, re-derivable audit chain ending in a single final hash. Nothing to bolt on later.
01b · Loading is not the bottleneck

A 1.24 GB book. Ingested, rebuilt, solved, validated — under a minute.

The slowest part of most rebalancing pipelines is getting the book in. We measured every leg on a real uploaded book of 250,000 accounts × 1,000 names, generated at full lot-ledger intricacy.

56s
server total — ingest + rebuild every account's tax economics + solve the whole book
28s
of that is the GPU solve (112µs/account); ingest 11s, tax-term build 17s
20s
to re-run the entire book for an intraday what-if — warm rescan
4.9h
the open-source reference solver's own time on the same book — the solve alone
Every leg is reported, not hidden. Each result breaks out ingest, per-account build, and solve time separately, so you see the whole cost — not a cherry-picked optimizer number. Validated in the same job: max weight error 0.0012% vs an independent open-source reference at 1e-9 tolerance. Measured 2026-07-13 on a datacenter GPU.
02 · Quick look

One POST. A job ID back. Poll for the result.

The full contract — datasets, polling, results, weights — is in the API docs. Here's the shape of it.

bash · submit a book-rebalance job
curl -X POST https://asymmetrycomputing.com/api/prism/v1/jobs \
  -H "X-API-Key: $KEY" -H "Content-Type: application/json" \
  -d '{"task":"book_rebalance","source":"stress_demo","n_accounts":25000,"n_names":1000,"validate_sample":50}'
03 · The economics

You don't buy GPU hours. You buy decisions.

The instinct that GPU compute is the expensive way to run optimization gets the unit wrong. An hourly rate only matters multiplied by hours — and PRISM doesn't use hours.

3¢
of GPU time to rebalance the full 500,000-account book — 47 seconds on one datacenter card at list price
~60×
smaller compute bill than the open-source CPU reference for the same book (~9 CPU-core-hours), at published cloud rates
4–27×
lower cost per solve than tuned CPU incumbents, measured across our cross-sector benchmark runs at real cloud prices
$0
commercial solver licences — the six-figure line item this API replaces outright
The objection, answered. Yes — a datacenter GPU rents for several times a CPU's hourly rate. It also finishes the same book hundreds of times sooner, so the bill is measured in seconds, not hours: at 94µs per account, a full rebalance costs a few millionths of a cent per account. The hourly rate is the wrong denominator. The decision is the unit that matters — and each one carries basis points of tax alpha, released capital, or captured dispatch revenue against a fraction of a cent of compute.
04 · Get access

Get a free trial key.

Trial keys are free, manually approved, and rate-limited — up to 25,000 accounts per job, 10 jobs per day. Enough to test a genuine book against your own numbers. Tell us who you are and what you're testing; we'll email a key.

✓Free trial, no card required.
✓Manually approved and emailed — not an instant self-serve key.
✓Up to 25,000 accounts per job, 10 jobs per day.
✓Full endpoint access — demo data, your own datasets, or a stress-test book.

Trial keys are free, manually approved, and rate-limited (up to 25,000 accounts per job, 10 jobs/day).

Read the full API reference.

Endpoints, job specs, dataset schemas, limits, and error semantics — everything you need to wire PRISM into your own pipeline.

Read the API docs →