Series · 6 posts · about 70 min

FinChat build log

Building FinChat, a reference retail-banking data and AI platform on Google Cloud: grounding, evals, policy as code, safety alerting, and ServiceNow integration.

  1. One source of truth for AI-agent grounding: OKF in practice

    July 1, 2026 · 6 min read

    I kept three copies of what my analyst agent was allowed to touch, and they drifted. Here's how the Open Knowledge Format cut it to one git-versioned source, compiled once and guarded by CI.

  2. Ontologies for mere mortals: one that actually runs

    July 12, 2026 · 11 min read

    Every ontology article leaves you more confused than when you started. Here is the opposite - one real ontology running inside a banking AI, every piece of jargon translated (taxonomy included), the drift it killed, and what it becomes at enterprise scale.

  3. Routing AI guardrail violations to ServiceNow

    September 4, 2026 · 7 min read

    Model Armor blocks a prompt and tells the user, and on most builds that is where it stops. Here is the full path from a guardrail violation to a correlated ServiceNow incident on Google Cloud, with screenshots from the running system.

  4. Policy as Code Without a Cluster: Adding OPA to FinChat

    September 5, 2026 · 16 min read

    FinChat already had five layers of policy, and every one of them evaluates an action that is already happening. None could see a change that does not exist yet. What OPA is, and how I filled that gap.

  5. Multi-turn context safety alerting: catching what a per-message filter cannot

    September 20, 2026 · 25 min read

    Per-message filters like Model Armor correctly pass most of a multi-turn probe. Adding the conversation tier that catches it, and the alerting fallback that turned out to matter more.

  6. Which parts of your eval framework you can stop maintaining

    September 27, 2026 · 5 min read

    A managed evaluation service can take over the judge, but not the contract tests, the drift canary, or anything needing provenance for the judge itself. Testing one produced four faults that each returned a plausible number instead of an error.