Project deep dives
These posts sit under their project pages and explain architecture, implementation, deployment, and lessons learned.
Why UK company due diligence is a workflow and evidence problem before it is an AI problem, and where an agentic research system can create practical value without replacing underwriting judgement.
How PolicyTrace handles disagreement between source documents with deterministic authority rules, conflict entries, evidence-backed review, and reviewer-visible arbitration.
How PolicyTrace turns extracted insurance fields into reviewable evidence with hidden field citations, Docling geometry, provenance matching, PDF highlights, and reviewer-visible source context.
How PolicyTrace designs prompts as one bounded layer in a reviewable Document AI system: document classification, specialist extraction, typed outputs, field citations, and evaluation gates.
How I would evaluate PolicyTrace beyond simple accuracy: golden examples, conflict fixtures, provenance checks, review outcomes, regression gates, and operational metrics.
A production hardening roadmap for PolicyTrace: authentication, durable storage, background jobs, audit trails, evaluation, monitoring, retention, and operational ownership.
How PolicyTrace packages a FastAPI backend and React review UI into one Docker app for a Hugging Face Space, and what still separates a public demo from production deployment.
How PolicyTrace turns extraction into a reviewer workflow with split-screen PDF evidence, field highlighting, confidence signals, verification, flags, and overrides.
How PolicyTrace turns overlapping insurance PDFs into a reviewable Golden Record by separating extraction from arbitration, conflict handling, evidence, and human review.
A layered architecture map for PolicyTrace, showing where parsing, privacy, model calls, evidence artifacts, arbitration, and human review belong.
