I'm a delivery manager with 18+ years in enterprise tech. I built SoWScanner to solve a problem I've had for years: vendor Statements of Work that look professional but contain vague deliverables, missing change control, and scope gaps that blow budgets.

The architecture separates two concerns:

1. AI extraction (Gemini 3 Flash) — reads the SOW and extracts structured factual observations: what's present, what's missing, what's vague. No scoring, no judgement.

2. Deterministic Python scoring engine — takes the extraction output and grades it across 8 categories using calibrated decision trees. No AI in the scoring loop at all.

Why the separation? Because LLMs are terrible at consistent scoring. They inflate scores, change their minds between runs, and can't explain their reasoning traceably. The deterministic engine solves all three problems — same input always produces the same score, and every score can be traced back to specific criteria.

Stack: React 19 frontend, FastAPI backend, Google Cloud Run, Cloudflare. Analysis takes about 30 seconds.

Free, no sign-up, documents are never stored: sowscanner.com

Happy to discuss the architecture, the scoring methodology, or the trade-offs involved.

The AI extraction + deterministic scoring separation is exactly right, and the reasoning holds beyond SOW analysis. I hit the same wall building internal decision tools: LLMs are useful architects but terrible engines. Same principle — AI reads and structures, deterministic compiler owns the output. Same input, same result, traceable, auditable. The governance argument alone justifies the split even before you consider reproducibility. What does your scoring engine do when the AI extraction is ambiguous or incomplete — does it fail loudly or degrade gracefully?