AI-assisted requirements review platform
Platform ownership and delivery of a document-to-review workflow, combining Python services, model integration and a browser workbench.
Context & responsibility
An internal R&D team was exploring how to connect engineering requirements with regulatory obligations while preserving human review and traceability. I owned the platform foundation and contributed across the workflow.
The delivered workflow
- 01Import Markdown / PDF
- 02Model-assisted mapping
- 03Human review
- 04CSV / Excel export
My engineering contributions
Platform & persistence
Bootstrapped FastAPI, configuration, typed domain models and CI. Implemented isolated project directories, atomic JSON writes, file locking and project lifecycle operations.
Review & auditability
Implemented review-state transitions with mandatory reasoning and optimistic concurrency control. Conflicting writes return HTTP 409. Contributed hash-chained audit logging and concurrency hardening.
Workbench & document flow
Built the browser workbench for importing documents, reviewing proposals and exporting results. Added Markdown/PDF parsing, content-based deduplication, CSV and styled Excel export with formula-injection protection.
Model integration
Implemented typed inference and health endpoints, plus automation for an Ollama service on remote GPU infrastructure through an SSH tunnel. Demonstrated a real inference call to the team.
Technical collaboration
Acted as PR reviewer, co-owned architecture decisions, shaped implementation proposals and temporarily facilitated Shape Up. Contributed across modules when the team needed additional coverage.
Outcome & scope
Demonstrated an end-to-end vertical slice that the team could extend: import, mapping, human review and export, backed by an extensive test suite of more than 800 automated tests across the project.