Portfolio Projects

Human-in-the-loop AI for financial workflows. A LangGraph agent drafts an auto-repair triage and estimate — calling tools over MCP for parts pricing and vehicle history — then returns a proposal, writing nothing to the system of record. A service advisor reviews in Vue and approves, edits, or rejects; only on approval does the C#/.NET backend commit the repair order and write an append-only audit record. The approval gate and audit trail are enforced server-side, with a test proving the agent cannot commit unilaterally. The whole request is traced end-to-end with OpenTelemetry — one trace from the browser click to the Claude call. One thesis, applied twice: put deterministic guardrails around a non-deterministic model. At runtime, a human-approval gate and audit trail wrap an LLM that writes estimates. At build time, hooks and CI wrap an LLM that writes code — three tiers of guardrails (intent, reflexes, law) that get harder to bypass at each level. A focused vertical slice (Customers, Vehicles, Repair Orders) on synthetic data, with auth, validation, and tests concentrated where bugs are expensive.