Traceable construction reporting

Problem

Construction teams needed one reporting model across daily-report PDFs, SharePoint, and FileCloud without losing tenant boundaries or document evidence.

Constraints

• Owners, contractors, and subcontractors required different project scopes.
• Extracted values needed to remain traceable to the original document.
• The platform spanned several repositories and worker processes.

My role

Led architecture and OCR work across FastAPI services, React workflows, PDF workers, shared Pydantic schemas, benchmarks, and reporting modules.

Decisions

• Kept raw and normalized data as separate records.
• Enforced project and tenant boundaries through scoped credentials and database policies.
• Reduced requirements to user behavior, domain invariants, failure semantics, and acceptance criteria.

What failed

• Early AI-assisted requirements were too verbose and prescribed implementation details.
• Engineers had difficulty separating required behavior from suggested implementation.

Tests and checks

• Used source-evidence contracts and real-document regression tests.
• Required idempotent worker completion and bounded retry behavior.
• Benchmarked extraction against retained document evidence.

What I would change today

Start with user behavior, domain rules, failure states, and acceptance criteria. Keep framework and service choices in architecture records after those requirements are clear.