An evidence-centric document intelligence platform for policy review, structured extraction, validation, and accelerated coauthoring.
Impact
policy review and coauthoring turnaround
Hiring signal
Document intelligence architecture, constrained LLM extraction, evidence traceability, deterministic validation, business-rule execution, and workflow acceleration.
My role
Lead developer · PDF normalization, evidence construction, hybrid retrieval, schema-constrained extraction, deterministic validation, backend integration, and audit artifacts
Status
An interactive trace is available.
The problem
Policy review and coauthoring require reliable extraction from long, structurally inconsistent documents containing mixed sections, tables, references, footers, and business-rule exceptions. A generic PDF chat interface is insufficient because reviewers need structured fields, source-linked evidence, validation, and reproducible outputs.
My approach
PolicyLens uses a staged processing architecture that normalizes PDFs through multiple parsers, anchors sections and tables, constructs evidence bundles, retrieves relevant context, performs schema-constrained LLM extraction, and applies deterministic Python rules for counts, scoring, invariants, and export-ready outputs.
Architecture judgment
The language model is restricted to retrieved evidence, while deterministic Python logic owns calculations, classifications, and invariant checks. This separation improves reviewability, repeatability, and operational suitability compared with an unconstrained PDF chatbot.
Architecture
The complete execution path, service boundaries, data stores, security controls, validation stages, and operational telemetry.
Implementation
Technology
Interactive trace
Run the trace to see the main workflow stages.