All earlier work ↗
Historical engagement record
About this record
This is an earlier engagement or reference design, preserved as authored. Reported metrics and delivery status have not been independently revalidated for this preview; source dates are not established here. Modeled figures and design targets are not realized results. This record does not substantiate outcomes for the proposed product.
Explore Tensile today ↗
Speed is easy to demo and dangerous to ship in a regulated finance function. The constraint is generating analysis you can put a senior name on. So the eval layer is the product: every run is checked for accuracy, PII-safety and authorization before it counts, and an independent reviewer model plays adversary against the first. The analyst still drives, queries and signs off in-document. Raw P&L in, senior-approved analysis out, with the governance wired through the middle.
The challenge
Analysts produce analysis documents from raw P&Ls inside a permissioned, collaborative hierarchy. They needed AI speed without losing governance: every output had to be accurate, free of PII leakage, and authorized for the person requesting it. In a regulated function, an ungoverned generation is a liability.
The approach
Retrieval from the firm’s approved data over the firm’s data and P&Ls feeding pre-vetted analysis templates, with multi-model agentic generation (Claude primary, routed per use case).
An eval trust-layer that enforces accuracy, PII-safety and enterprise authorization on every single run, as it happens.
A maker-checker adversarial review where an independent model flags inconsistencies before sign-off, plus in-document highlight → query → resolve and senior approval inside permissioned spaces.
Artifacts and handoff
A built financial-analysis harness: governed-context RAG, templated multi-model generation, the per-run eval trust-layer, the adversarial maker-checker reviewer, and the in-document review/approval workflow.
Reported measures and design targets
Eval gate on every run: accuracy, PII-safety and enterprise authorization enforced per generation
Maker-checker built in: an independent model flags inconsistencies before sign-off
Retrieval from the firm’s approved data: generation grounded on the firm’s own data and P&Ls
Multi-model routing: Claude primary, routed per use case
Explore the failure mechanisms, the tests they need and the evidence that establishes the customer outcome.
AI reliability, from production to customer outcome.