Academic Commercialization Assessment Agent
An evidence-constrained agent that assesses the commercialization potential of research topics: deterministic retrieval, a six-stage LLM workflow, and an auditable, cited scorecard.
Evidence shows deliberate privilege design: subprocess isolation, write-once terminal records, 128-bit run URLs as read capabilities with explicit do-not-publish warnings, admin-code-gated mutations, paid-operation quotas. Deductions: ownerless BYOK runs have no second server-side identity; dependency security rests on version constraints (crewai pinned, others ranged) with no vulnerability-audit or lockfile evidence in scope.
Pydantic contracts, deterministic weighted scoring, bounded reviewer corrections, a zero-provider test suite with an 85% coverage floor support self-consistency; conftest systematically strips real credentials and access codes to prevent leaks, and offline degraded paths are explicitly tested. Deductions: user-facing failure message text is not directly visible in the provided source; the non-exactly-once nature of recovery is asserted only in README prose.
Capability boundaries are exemplary: the README states reports are not legal/investment/FTO due diligence, marks experimental features as disconnected from production, documents single-replica scale and benchmark gaps (multilingual, short, non-technical). Deductions: trigger/routing behavior (LLM_PROVIDER auto-selection order, Tavily precedence) is prose-only with no code-level evidence; audience is defined but scenario coverage rests on a 10-topic technical benchmark.
Good information architecture (docs map separating current guides from dated decisions), complete quick start (uv, .env, CI matrix, Docker), stable naming (2.0.0, fixed entry points), full MIT license, maintenance supported by CI and contribution rules. Deductions: no CHANGELOG visible, no FAQ, and AGENTS.md plus several docs files are referenced but not present in evidence, so their claims cannot be verified.
Output is usable (Markdown/PDF export, auditable scorecard, recovery children, shareable links); cost accounting is thorough (complete/lower-bound/unavailable states, quotas, ablation showing the four-node arm cheaper). Deduction: marginal value is self-weakened by the project's own evidence — the five-reviewer utility study missed its registered success rule and both target-user pilots answered MAYBE on reuse without checking external sources, so adoption value is unproven.
Fact/inference separation is the repository's strongest suit: the README systematically distinguishes measured results from unvalidated claims (TRL ranges adjusted post-hoc so not held-out, the uncited-numeric proxy misses hallucinations, 30/30 recovery is not an exactly-once guarantee), tests strictly separate not_checked from fail statuses, and valid citation IDs are explicitly disclaimed as not entailment. Deduction: traceability depends on runtime artifacts and an evidence ledger referenced but not provided, which static review cannot inspect.
- Static review with low confidence: no runs executed; 30/30 completion, TRL calibration and recovery behavior were not independently verified.
- Anyone holding a full run URL can read that run; configure access control, paid-operation limits and retention before public deployment.
- Ownerless BYOK runs have no second server-side identity — the URL is the sole capability, and leaking it is unrecoverable.
- Deploy with one application replica / one Uvicorn worker: in-memory ownership and file-backed quotas are not a distributed queue.
- Tool-calling v8 failed three of six gates on the unseen set and remains disconnected from production; do not tune on the failed set and integrate.
- AGENTS.md, the evidence ledger and most docs files were not provided in this evidence set; their contents are unverified.
- Real analysis incurs provider cost; durations of 306–885 seconds are observed, not a promised SLA.
What does this agent do, and when should you use it?
This open-source project (GitHub: shuxiachai/academic-commercialization-agent) assesses how far a research topic could travel toward commercialization, from technical maturity to market signals. It first runs deterministic retrieval over academic, patent and market sources, validates them into a frozen source registry, then runs three parallel evidence specialists (Academic, Patent, Market) followed by Writer, Reviewer and Scorer stages, with Scorer applying a deterministic weighted formula. The stack is Python, CrewAI, FastAPI and a build-free vanilla JavaScript client, exposed through browser, CLI and HTTP interfaces, deployable via Docker/Railway. Autonomy is deliberately limited: retrieval is not free-form, every claim carries citation IDs and reliability warnings, and reports export to Markdown or PDF. An interrupted run recovers as an immutable child from its longest validated checkpoint prefix. The authors publish a frozen baseline of 10 topics × 3 live repetitions and state plainly that reports support research triage only, not technical, legal or investment due diligence.
You submit a research topic or attach a paper PDF, optionally with Decision Context, which lets the system distinguish exploratory assessment from an actor-specific decision and whether a threshold is owner-approved. The flow is: deterministic retrieval (source-native clients plus Tavily/Serper web search) → URL/DOI checks, provenance tiers, deduplication and registered source IDs → frozen source registry → parallel Academic/Patent/Market specialists → Writer → bounded Reviewer corrections → Scorer with a deterministic weighted total. Output is a cited report with reliability warnings and a scorecard, exportable as Markdown or PDF and shareable via a run link. Run artifacts persist through content-addressed checkpoints; interrupted runs recover as immutable children. Deployment is a FastAPI application (one replica / one Uvicorn worker) with a browser UI, plus a CLI: uv run academic_agent --topic "solid-state batteries for electric vehicles". Run URLs carry 128 bits of randomness and act as read capabilities; mutation additionally requires an operator code or BYOK credentials.
- A tech-transfer officer with a paper or topic direction needs a cited preliminary TRL and market-signal read before pursuing engagement.
- A research group wants to cheaply triage out clearly immature topics before paying for commercial consulting on a translation grant application.
- A patent analyst wants academic literature, patents and market evidence aggregated in parallel as groundwork for deeper due diligence.
- An investor or incubator uses the Decision Context feature to anchor an assessment to a specific threshold and decision-maker rather than a generic feasibility report.
- An engineering team studying constrained multi-agent architectures can study its deterministic retrieval, frozen evidence registry, checkpoint recovery and published measurements.
What are this agent's strengths and limitations?
- Evidence constraint is a real differentiator: retrieval is deterministic, claims bind to registered source IDs, and the frozen baseline showed 0 unsupported numeric lines across 30 reports.
- Strong auditability and reliability engineering: content-addressed checkpoints, immutable recovery children, write-once terminal records, and 30/30 offline fault-injection recovery completions.
- Multi-provider support (Qwen/DeepSeek/OpenAI/Anthropic) plus BYOK mode avoids single-vendor lock-in; the browser client supports bring-your-own-key.
- Visible engineering quality: a 2071-test / 678-subtest baseline, CI across Linux/Windows × Python 3.11/3.12, an 85% coverage floor, Docker and zero-provider smoke tests.
- Honest evidence disclosure: a 90-cell topology ablation, reviewer utility study and user pilot are published with limitations, including failed gates.
- Cost and latency are nontrivial: real analysis consumes paid provider quota, single Qwen completions took 306–885 seconds, and PDF extraction is a paid operation too.
- Hard deployment constraints: in-memory ownership and file-backed quotas force one application replica / one Uvicorn worker — not a distributed queue — so scaling requires rework.
- Limited evidence coverage: TRL calibration was 26/30 with ranges adjusted after early observations, not independent held-out validation; the 10-topic benchmark excludes multilingual, short or non-technical topics.
- Weak product validation: a two-person target-user pilot retained DEFER with MAYBE on reuse, and the reviewer utility study missed its registered success rule — adoption value is unproven.
- Reports support research triage only, explicitly not technical, legal, regulatory, investment or FTO due diligence; valid citation IDs do not prove a source entails a claim.
How do you install or deploy this agent?
Requires Python 3.11 or 3.12 and uv: git clone https://github.com/shuxiachai/academic-commercialization-agent.git && cd academic-commercialization-agent && uv sync. Copy .env.example to .env and fill in credentials, e.g. for Qwen: LLM_PROVIDER=qwen, DASHSCOPE_API_KEY=your-key, QWEN_MODEL=qwen3.5-plus, QWEN_API_BASE=https://dashscope.aliyuncs.com/compatible-mode/v1, TAVILY_API_KEY=your-search-key. With multiple LLM keys present, set LLM_PROVIDER explicitly; otherwise auto-selection is DeepSeek → Qwen → Anthropic → OpenAI. Tavily takes precedence over Serper when both search keys are present. Before public deployment, configure access control, paid-operation limits, retention and persistent storage, and keep one application replica.
How do you use this agent?
Start locally with uv run uvicorn api.main:app --reload, open http://localhost:8000, then submit a topic or PDF and choose report language and scoring profile. CLI alternative: uv run academic_agent --topic "solid-state batteries for electric vehicles". You can also call the FastAPI HTTP endpoints documented in docs/operating-guide.md. Follow progress, citations and reliability warnings live; export Markdown or PDF and share the run link when done. If a run is interrupted, recover it as an immutable child using its longest validated checkpoint prefix and fresh credentials. Note that real analysis incurs provider usage; observed Qwen completions took 306–885 seconds, and the project promises no three-minute SLA.
How does this agent compare with similar options?
The project names no direct competitor. Its positioning differs from general autonomous agent frameworks (such as CrewAI, which this project uses as an orchestration library): the differentiation is deliberately compressed autonomy, with deterministic retrieval and auditable scoring replacing free-form tool calling; its Tool Calling extension remains in zero-call shadow mode and is not authorized for production.