// the find
shuxiachai/academic-commercialization-agent
Turn papers and research topics into source-linked commercialization assessment drafts. Python, FastAPI and CrewAI; auditable scoring, checkpoint recovery and durable request receipts.
A FastAPI/CrewAI pipeline that takes a research paper or topic and produces a commercialization assessment: parallel academic/patent/market evidence gathering feeding into a writer, reviewer, and deterministic scorer. Aimed at people doing research triage (tech transfer offices, VCs scanning academic output) who want a repeatable, source-linked first pass rather than a finished due-diligence report.
Treats LLM output as something to constrain, not trust: Pydantic contracts, deterministic weighted scoring, bounded reviewer corrections, and citation/claim screens that distinguish 'unavailable' from 'passed' rather than silently skipping. Recovery is genuinely engineered — content-addressed checkpoints, subprocess isolation, immutable recovery children, and durable receipts so a paid LLM call isn't lost on a dropped connection. Test coverage is real (2071 tests, Linux/Windows x Python 3.11/3.12 CI, 85% coverage floor) and they run actual ablations (4-node vs 6-node topology, cost/token deltas) instead of just claiming multi-agent is better.
The docs directory is dozens of dated prereg/results/errata files that read like a raw experiment log checked into git rather than curated documentation — there's no clear entry point for 'how does scoring actually work' without reading the changelog of every self-correction. TRL calibration lands at 26/30 on a 10-topic x 3-run benchmark where the expected ranges were adjusted after early observations, so it's closer to fitting your own test than independent validation. The 'multi-agent' and tool-calling story oversells what ships: supplementary retrieval is shadow-mode with zero live calls, and the one production tool-use feature is a single bounded selection over already-saved text, not autonomous research. Explicitly single-replica with in-memory state and file-backed quotas, so it won't survive a second worker without a rewrite of ownership/quota handling.