Intelligence that can reason.
Infrastructure you can trust.
TSIQ turns data from TS Imagine's products — TradeSmart, RiskSmart, PrimeOne — into actionable, explainable intelligence. One client-facing agent, DATA_ANALYST, answers questions and runs tasks across the desk: slippage and execution quality on the trading side, exposure and stress explanation on the risk side, positions, balances and cash flows on the financing side.
Underneath is a two-plane architecture. The probabilistic plane proposes — outputs typed and disposable, nothing decisive. The deterministic plane is the record — exact, sequenced, replayable. A typed seam connects them: semantic models, certified queries, entitlements, evaluations, evidence and audit.
The model never bypasses TS Imagine's entitlements, controls, calculations or execution logic. Nothing crosses the seam untyped — the AI can never corrupt the record.
COVERAGE Trading & liquidity analytics and product usage live today — risk and securities financing coming online through PS1, the portfolio-state database.
Three classes of action. Three levels of control.
Where the AI stops and governed financial infrastructure begins is not a gray area. Every request is classified before anything happens, and the class determines the controls applied.
Examples
- Investigate where the slippage came from on this order
- Explain what moved this portfolio's exposure and stress results
- Answer a question about positions, balances or financing cash flows
Control applied
AI may respond directly. Answers are entitlement-filtered, grounded in deterministic calculations, and delivered with evidence and lineage attached.
Two planes. One seam.
The probabilistic plane proposes. The deterministic plane is the record. The seam is the typed contract between them — nothing crosses untyped. Click any component to light up everything it talks to — or run the trace to watch a real request move through the system.
Click a node to expand it and see a real example of what it does · click the background to clear · amber = probabilistic · teal = the seam · blue = deterministic
Not drawn as nodes because it touches all of them — the same identity gates the data and the agents.
Data
Five databases — DM1 · TCLA · PS1 · Pulse · Taia — behind a three-zone ELT pipeline: RAW → STAGING → OPTIMIZED (gold · 180 tables), with column-level lineage-as-code.
Compute
The TS Imagine Master Snowflake Account (GCP, US) feeds one governed TSIQ Account per client. OneTick computes tick benchmarks at the speed of the source.
Trust
One identity gates the data and the agents — SSO + Zscaler + MFA + RBAC. Per-account MCPs are inactive by default and expose only their own account's data.
Your AI. Our backbone.
TSIQ meets clients wherever their AI lives — desktop agents, CoWork, AI browsers, or natively inside Snowflake — through one master MCP. Behind it sit other MCPs or plain REST calls into TradeSmart and RiskSmart — an implementation detail, never a second door. And it runs both ways: Cortex Code and CoWork can call out to any MCP a client runs — their own, Atlassian, anyone's. Click any component to light up its path.
Roundtrip by design · TradeSmart → TSIQ → TradeSmart · every TSIQ surface is HTML5 — an AI browser can read and operate it directly · per-account isolation, one identity
Nothing crosses untyped.
Every crossing between the probabilistic plane and the record carries the same seven things. No exceptions, no side doors.
Who is asking — verified, not asserted. SSO + MFA + RBAC.
Which account, desk, order and session — canonical IDs, resolved.
What this user is entitled to see and do.
Structured {sql · narrative · sources} — never free text. Op-IDs resolved before honored.
Access and quality rules enforced as code.
Benchmark Audit JSON — every number explainable.
An append-only, replayable record of the exchange.
What's replaceable — and what isn't.
This is not "Snowflake plus an LLM." The ecosystem components below the line are commodities by design. The seam and the record took five years and $100M to build.
Swappable by design
Best-of-breed today, replaced tomorrow without touching the seam.
- Foundation models — Claude, OpenAI, Mistral, Llama, Gemini
- AI UIs — Claude Desktop, ChatGPT, CoWork
- AI browsers
- Dashboards — build your own
- Authoring — Cortex Code, Snowflake CoWork
- Embedding models & vector stores
- Model hosting
The seam & the record
The governed financial substance no model provider can replicate.
- The TSIQ Ontology · 12 semantic views
- 1,700+ prompts with reference answers
- Best Execution Engine
- OPTIMIZED gold data with full lineage
- Durable Op-IDs & canonical identifiers
- Benchmark Audit JSON transparency
- Entitlement model — one identity, data and agents
- Certified, continuously-evaluated agent behavior
- 5 years · $100M · a 40+ person data science team
Pay for intelligence once.
The industry has finally named the problem. Enterprises using closed frontier models "pay for intelligence twice" — once with money, and again with the proprietary knowledge they must reveal to make the model useful. The prompts people write, the workflows they build, the corrections they make when the model is wrong — the exhaust — leak into the vendor's learning loop, trace by trace, eval by eval. Learning flows in one direction, and value converges toward whoever owns the learning infrastructure — not toward the people whose knowledge it is.
The prescription now on the table is a real trust boundary: a perimeter where an enterprise's data, traces, evals and memory compound together — and nothing, exhaust included, leaves without consent. Five demands follow from it. TSIQ's architecture reads like it was designed against the checklist — because, in effect, it was.
Own your evals, your traces, your organizational memory — evals define what "good" means for you.
1,700+ prompts with reference answers, Benchmark Audit JSON, append-only audit — accumulating in the client's account, owned by the client.
A learning environment inside your own tenant, where models work real workflows without exposing your knowledge.
Claude and five other providers run inside the client's Snowflake account. DATA_ANALYST works the desk's real workflows — and nothing crosses the perimeter.
Decouple orchestration from any single model — if one is taken away, can you still operate?
The seam is that decoupled layer. Take any model away and the certified behavior, ontology and evals — the veteran capability — remain intact.
Bring context, models and tasks together efficiently, without sacrificing quality.
The right model per task across six providers — and certified queries answer known questions the vetted way, without burning tokens re-deriving them.
Close the loop — a continuous learning machine that compounds your firm's value.
Every question, correction and evaluation feeds the Believability Loop. The hill-climbing machine belongs to the client — not to the vendor.
Product vs. model
What happens when the model sits outside your perimeter.
- Prompts and workflows leave the building
- Corrections distill into the vendor's model
- Your evals build someone else's know-how
- The better it works, the more you reveal
- The vendor can end up competing with you
Product with model
Claude and five other providers run inside the Snowflake perimeter — in the client's own governed TSIQ account.
- Prompts, corrections and evals never leave the perimeter
- The exhaust feeds the Believability Loop — for the client
- 1,700+ prompts accumulate as the client's asset
- One identity, one boundary — SSO + MFA + RBAC
- Models stay swappable; knowledge stays put
- One payment. The second one never happens.
PRODUCT WITH MODEL — NOT PRODUCT VS. MODEL. In consuming intelligence you create intelligence — and inside TSIQ's perimeter, what you create stays yours.
A note on tooling. TSIQ's artifacts — dashboards, reports, evaluations, and pages like this one — are largely authored with Cortex Code operating inside the same governed Snowflake environment: the same identity, entitlements and audit that gate production data gated the build. It's a useful proof point about development velocity. It is deliberately not part of the architecture story above, because how an artifact was coded matters far less than how the production system is controlled.
TS IMAGINE · TSIQ · ARCHITECTURE BRIEFING · 2026
Book a TSIQ demo
Bring your questions — or just ask us what TSIQ can do.
Keep reading
