Six AI agents grade your lab work for market readiness

What it does

Feed it a research topic and six CrewAI agents fan out in parallel to gather academic papers, patent filings, and market intelligence. They synthesize the evidence into a structured commercialization report with inline citations, a quantitative scorecard, and TRL/MRL ratings. The result is essentially a VC diligence memo generated from live sources.

The interesting bit

The scoring agent consumes the raw evidence JSON directly, not the polished report text, so its ratings are insulated from the writer’s prose. A reviewer agent guards against unsupported numeric claims and citation drift, and the entire pipeline runs in an isolated subprocess so a run can be terminated immediately without waiting for the current LLM call to finish.

Key highlights

• Six specialized agents (academic, patent, market, writer, reviewer, scorer) with parallel evidence gathering
• Inline citations ([A1], [P2], [M3]) mapped to validated sources from OpenAlex, Semantic Scholar, Google Patents, and domain-allowlisted market reports
• Reproducible temperature=0 JSON-mode scoring with traceable source IDs for each dimension
• Live demo on Railway with bring-your-own-key or access-code entry; keys stay in the run’s subprocess environment and never touch disk
• Multilingual input handling with fully localized reports in Chinese, Japanese, Korean, German, French, and others

Caveats

• TRL calibration ranges were revised after earlier scores were observed, so the benchmark shows consistency rather than blinded accuracy
• Repeat runs on the same topic yield close but not identical scores because live market search results shift between sessions
• Patent search relies on Google Patents and WIPO only, since the USPTO PatentsView API was retired

Verdict

A solid fit for tech transfer offices, academic VCs, or research groups that need fast, structured landscape assessments. Not a replacement for legal freedom-to-operate analysis or fully deterministic benchmarking.