Why a PM should care how it is built: the whole book is repriced overnight on one calculator, the agents argue over those same numbers, and every figure opens to the filing and the tick that produced it. Below — the gateway, the typed bitemporal store, the agent committee and the audit trail, for the people who want to see it.
Every request enters through one gateway and fans out to isolated services — each one can fail without taking the platform down.
A failed non-critical router returns 404; the rest of the platform keeps serving. Only auth, admin and datasource-health are critical — their failure deregisters the replica.
Every router is registered through a single manifest — module · prefix · deps · critical · node. No second include path. Adding a domain is one manifest entry.
Stateless math nodes scale to N workers; rate-limited feeds and stream consumers stay pinned to 1×. Isolated services behind one gateway.
Light calls run straight through a strict layering. Anything CPU-heavy returns a job id immediately and computes on a worker — the user never waits on a blocked request.
Router → Service → Repository. Database calls live only in the repository layer — no SQL leaks into business logic.
A stress or backtest call returns 202 + job_id instantly. The client polls a job hash; the worker owns the long-running math.
Every financial model rejects NaN/Inf at the boundary. ValueError → 400, payloads capped, jobs delivered at-least-once.
Every proxied node sits behind a circuit breaker. A slow or failing node trips, stops taking traffic, and probes its way back — without a human paging in.
An asyncio task per replica polls each node's /health every 5s with 3-layer jitter. Liveness and circuit state stay independent.
OpenTelemetry spans on every request. 10% base sample, 1% on health checks, 100% on 5xx. An X-Trace-Id on every response.
Feature flags with a 60s cache. 3 SDK failures trip a 30s circuit to a safe default — flag-service downtime can never add latency.
Market data, alt-data and compute analytics flow into a single bitemporal store. Agents query typed objects — not SQL.
Ask "top-10 option positions by contribution to portfolio Vega, where the vol surface is older than 30 minutes" — and a normal stack runs ad-hoc SQL across five tables with no guarantee the joins are even semantically correct. Multiply that by every desk, every regulator question, every replay request. The ontology removes that whole class of question from the human queue: every computed object is a typed first-class citizen, and the query is a typed traversal — not a SQL gamble.
Without this layer, AI agents cannot reason. With it, a multi-hop question — across positions, models, market data and stress results — answers in one traversal of one graph, with the time axis carried through every hop.
54 typed object kinds across 5 categories — financial seed, alt-data, risk, model, market data. Each typed kind is a roadmap card with its own shipping quarter. See ONT-01 → ONT-12 on the roadmap →
Every fact carries event-time (valid_from / valid_to) and system-time (recorded_at / superseded_at). An audit replay returns what the system knew at any past moment — not what it knows today.
Pydantic v2 discriminated union on kind. URN scheme nq:<kind>:<local-id>. ClickHouse ReplacingMergeTree as the materialised store. 49 typed mappers on the read path.
A @persist_to_ontology decorator wraps compute endpoints. ClickHouse outage degrades the graph, never the user request. Native hot-path tables stay authoritative.
The architectural bet. Every audit, every replay, every cross-desk query reads from one truth on one time axis. Building this layer first is what makes every other module Nyquist ships compound instead of fragment.
36 agents, each on a named public investing methodology and its published corpus. Every query runs as a debate cycle (claim, counter, tail-check, consensus), step-by-step logged. The orchestrator N.Y.Q.U.I.S.T. routes the swarm and recomputes every number through a deterministic calculator before delivery.
Each agent runs on a published corpus — books, public speeches, shareholder letters, papers. The lineage is shown in small caps. Three anchors below; all 36 in the roster.
Concentrated quality at fair price. Mental models. "Invert, always invert."
Macro-asymmetric bets. FX/rates/commodities-aware. Position size on conviction.
Convex tail-risk hedging. "What kills you when nothing's wrong?"
Agents are named for public investing methodologies; the lineage shown is a stylistic homage. Not affiliated with, endorsed by or connected to any named individual, their firms or their estates.
Every query runs a debate cycle: thesis from the quality lens, counter from the macro-asymmetric lens, veto from the convex-hedge lens. Output is consensus — or marked dissent, never a single unchallenged claim.
A backronym that is a contract. Every letter is a constraint in production code, not a marketing line. The orchestrator routes the 36-agent swarm, recomputes every number through a deterministic calculator, and emits an audit trail at a financial regulator's evidentiary grade.
Why it's defensible. Investing is a debate between frameworks. The orchestrator surfaces that debate — agreements, dissents, route taken — instead of hiding it behind one confident answer. Wins the IC meeting and the inspection.
Who decides — and who doesn't. The swarm briefs, argues, recomputes, logs. It never places an order, signs off on a number, or commits capital on its own. Treat it as a research desk on a senior PM's calendar — it does the homework, surfaces the dissent, drafts the memo. The decision, the signature and the accountability stay with the human operator.
A pure calculator recomputes every number before delivery. The model can draft and explain — it cannot invent a final number. What passes the CRO's bar passes the regulator's.
LLM call → schema validation → pure-function recompute → drift check → audit log. A language model cannot ship a risk number that hasn't been sanity-checked by deterministic code.
Domain SLM fine-tuned on 7,553 regulatory documents; runtime through NVIDIA NIM, deployable on-prem or in our managed cloud — not yet serving, see /roadmap. Whitelisted tool registry, enforcement before dispatch. CRO-grade permission model day one.
Thesis: AI becomes usable in regulated finance by making the workflow controllable, auditable and safe. Not by making the model smarter.
Full demo in your browser — no calls, no scheduling. Watch the agents debate, query the ontology, inspect every reasoning step.
