The firm was gated by manual case processing — medical-legal reports read by hand, case knowledge stuck in a few people's heads. I designed and shipped a production platform that reads the case medicine, routes the work, and forecasts settlements.

An end-to-end agentic pipeline: it ingests case documents, runs multi-step LLM analysis against the firm's own data with vector retrieval, extracts the structured findings (examining physician, date of injury, impairment rating, P&S status), auto-generates the memo, and routes it to the right staffer — every run cost-instrumented.

In production since 2025 and still running daily — hundreds of case documents processed, firm-wide. A comparable build commissioned today runs well into six figures to replace; this was delivered solo.

California workers'-comp permanent-disability rating is slow, error-prone, expert work. I built wcpdr: an AI-agentic calculator that produces a rating from the medical report — a product with no off-the-shelf equivalent.

Domain-modeled the rating rules, wrapped them in agentic logic over the report text, and shipped it as a live application with real firm users — product thinking, not a notebook.

In daily production use since September 2025 — 500+ cases processed for its client firm, about 40 a month, at under $3 of AI per case. I don't just integrate models; I turn hard domain problems into products people rely on.

Case knowledge was bottlenecked in a handful of people. I put a conversational assistant inside Microsoft Teams, available to every staffer, answering questions about live cases on demand.

Retrieval-grounded over the firm's own case corpus and wired into the Teams workflow staff already live in — so the answer comes to them, where they work, with no new app to learn.

In daily firm-wide use, a year on. Built in 2025, when getting LLMs to work reliably on real case data — not a demo — was genuinely hard.

New-case intake is a slow, people-heavy funnel. I built an agentic pipeline that takes an intake and produces the finished case package — standardized, logged, and costed per run.

Multi-step agents extract and structure the intake, assemble the package against the firm's templates, and hand it to staff for the human step — governed and traceable throughout.

Delivered 60 real case packages in production at roughly 13% of the equivalent staff cost — an ~87% cost reduction on that workflow.