Nyquist
CLOSED BETA · 2026
Start a 60-day pilot →
Nyquist
Most "AI for finance" is a chat window on top of someone else's data. We went the other way — down to the schema — so a PM gets three things: the whole book repriced overnight, a debated call with an invalidation trigger by the open, and any number opening to the filing and the tick that produced it. Underneath: a typed financial ontology, a single compute engine, and an audit trail that is a property of the architecture, not a feature on a roadmap.
Each layer is owned, so there are no brittle handoffs between vendors and no seam where data gets dropped on the floor. The agent layer can reach every layer beneath it.
★ Each layer is owned end-to-end — the source of the moat and the audit trail.
Not a data lake. Every object in the system is a typed entity with explicit references — a Trade points to its RFQ, its Instrument, its Counterparty and its Fund. That typing is what lets an agent reason about the book and a supervisor trust the number.
★ Illustrative layout, not live counts. Risk metrics: VaR, ES, vol, beta, Sharpe — typed & audited. Models versioned & signed. Reg corpus: Basel III, MiFID II, IFRS 9.
Pick a figure in any report and replay it. Below: a single VaR-99 delta, unwound to the two exchange ticks that priced the trade. Every node is content-hashed and signed — the same inputs reproduce the same number, every time.
Reproducibility isn't a report you generate — it's a property of the system. Every agent call, stress run and number is replayable to its source, because the lineage was never optional.
Agents call typed primitives through an MCP server and act on real state. The output is a report with a decision and an invalidation trigger — not a transcript you re-verify by hand.
On-prem or private cloud, with the model and compute inside your perimeter. The deployment posture institutional buyers — and their supervisors — already require.
Funds, banks and regulators share the ontology, the compute and the log. A feature for one operator hardens the product for the others — the codebase compounds.
Positions, filings and feeds are mapped onto the typed ontology. Lineage is established at ingest, so every downstream number already knows where it came from.
The swarm reads the corpus and the book, debates the thesis from multiple stances, and converges on a confidence-weighted position with an explicit invalidation trigger.
Pricing, risk, tail fits and scenarios run on the shared engine — the same numbers across research, the desk and a supervisor's review.
The output is a written report with a decision and a full audit log. Any number can be replayed to the filing that produced it.
The posture institutional buyers and supervisors require — not a SaaS that asks them to ship their book to someone else's cloud.
The full platform runs in your browser — a pre-loaded $120.4M book is wired in. Watch the agent swarm debate, run stress, and replay any number back to its filing. No calls, no scheduling.