v2.0.0 - Reliability-first production workflow

Academic Commercialization Assessment Agent v2.0.0

Reliability-first production workflow

Version 2.0.0 turns the original Gradio demonstration into a deployable,
evidence-constrained decision-support system. A FastAPI service, build-free web
client, CLI, and recovery path now observe the same persisted run artifacts.
The release also adds measured evaluation, paid-operation controls,
privacy-reduced tracing, and node-level crash recovery.

This remains decision support rather than legal, regulatory, technical, or
investment advice.

Highlights

Evidence before generation

• Deterministic retrieval plans free-form topics and collects academic, patent,
  market, and applicable regulator or trial-registry records before CrewAI runs.
• DOI/URL checks, allowlists, provenance tiers, relevance screening,
  deduplication, and a rejection audit produce a frozen source registry.
• The six-stage workflow analyses that registry, writes a cited report, applies
  a bounded review plan, and generates a source-traceable scorecard.
• Evidence-gap planning is available as zero-call shadow instrumentation. It
  records whether bounded supplemental search would be justified but performs
  no production Tool Calling in this release.

Public application and paid boundaries

• Gradio has been replaced by FastAPI plus a vanilla HTML/CSS/ES-module client.
• Runs execute in cancellable subprocesses and expose live progress, sources,
  scorecards, reliability states, usage, cost, and downloadable reports.
• Access-code and BYOK modes isolate credentials and run history.
• Complete runs and PDF extraction share concurrency and daily paid-operation
  admission; failed startup rolls back reserved quota.
• Rate limiting, ownership checks, bounded uploads, retention, URL validation,
  and strict response headers harden the public surface.

Auditable recovery and observability

• Content-addressed checkpoints bind reusable outputs to input, evidence,
  configuration, task, and pipeline identity.
• Recovery creates an immutable child, requires fresh credentials, validates
  the longest contiguous prefix, and exposes reuse separately from persistence.
• OpenTelemetry and OpenInference connect retrieval, CrewAI tasks, provider
  calls, and quality checks in one privacy-reduced trace. Local artifacts remain
  authoritative when the collector is unavailable.

Measured evidence

• Baseline: 30/30 end-to-end completions, 26/30 milestone-anchored TRL
  matches, 30/30 formula checks, 30/30 report-structure checks, and 0 unsupported
  numeric lines. This is a revised calibration baseline, not a held-out score.
• Topology ablation: 90 paid cells across 1-, 4-, and 6-node workflows. The
  4-node arm used 54.89% fewer median tokens and 47.03% lower median cost than
  the 6-node arm. This supports domain decomposition, not every production node.
• User utility: five reviewers completed 20/20 eligible blinded judgments.
  The full workflow won decision usefulness 6:4 in each round but failed the
  pre-registered success rule; the monolith led information gain 11:5.
• Fault injection: 30/30 zero-network recovery children completed, skipped
  90 committed task executions, and duplicated 0 task executions.
• Production recovery: one post-fix same-revision child reused a four-node
  prefix, made 0 new evidence-agent requests, and completed the suffix. The
  interrupted source's usage was unavailable, so total cost and general savings
  remain uninspectable.
• Verification: 1,391 tests plus 627 subtests, Linux/Windows × Python
  3.11/3.12 CI, and 87.01% measured coverage above an 85% floor.

The protocols, row-level artifacts, failed candidates, and caveats are linked
from the README
and summarized in the
portfolio case study.

Breaking changes from v1.0.0

• The Gradio UI and python app.py development path are retired. Start the web
  application with uv run uvicorn api.main:app --reload.
• The HTTP API and persisted run contract are now the supported integration
  surfaces. Consumers of the original UI callbacks must migrate to the FastAPI
  endpoints documented at /docs.
• Public deployment configuration now separates operator-funded access codes,
  BYOK credentials, paid-operation limits, retention, and optional telemetry.
  Review .env.example rather than carrying forward a v1 environment blindly.

Run locally

uv sync
uv run pytest -q
uv run --with ruff ruff check .
uv run uvicorn api.main:app --reload

Then open http://127.0.0.1:8000. Real analysis requires one supported LLM key
and one search-provider key; the test suite is zero-network.

Known limits

• The user-utility panel is small and mostly proxy users; it does not establish
  adoption, ROI, or better real-world investment decisions.
• File-backed run state and paid-operation accounting target one application
  replica. Horizontal scale requires a transactional shared store and queue.
• External provider requests remain at-least-once at interruption boundaries.
• Production supplemental Tool Calling is intentionally disabled until the
  shadow planner demonstrates useful evidence gain at acceptable error, cost,
  and latency.
• Code-package analysis is not implemented. Patent relevance has one human
  label set but no second independent reviewer or inter-rater agreement.
• Precision-first regulator-title recovery covers only exact FDA 510(k) and
  ClinicalTrials.gov URL shapes and is not evidence of title truth.

Portfolio material

• Live application
• One-page case study
• 90-second demo recording guide
• Full README

v1.0.0

Academic Commercialization Assessment Agent — v1.0.0

First public release of a multi-agent system that turns any research paper or
topic into a structured commercialization readiness report with verifiable citations.

Features

Multi-agent pipeline (6 agents, CrewAI)

• Agent 1–3 run in parallel: Academic Literature Analysis, Patent Landscape, Market Intelligence
• Agent 4: Report writing with inline citations [A1][P2][M3] and auto-retry guardrail
• Agent 5: Quality review with 6 citation integrity rules
• Agent 6: Quantitative scoring (TRL / MRL / Patent / Market / Evidence), formula auto-corrected

Source collection (deterministic, pre-run)

• Academic: OpenAlex + Semantic Scholar, DOI deduplication, Crossref citation backfill
• Patent: Serper → Google Patents / WIPO, URL reachability verified
• Market: Serper + 30+ domain allowlist, low-quality sites removed

Scoring

• Five dimensions with five industry weight profiles: industrial / biotech / material_science / software / deep_tech
• Each dimension cites supporting source IDs (traceable to validated sources)

Gradio web UI

• Live per-agent progress with elapsed time
• Radar chart + bar chart scorecard with source ID chips
• PDF upload: anchor the analysis to a specific paper (becomes source A1)
• Markdown report + PDF export (CJK font support)
• History tab with all past runs

Multilingual output

• 12 languages auto-detected from topic: English, Simplified/Traditional Chinese, Japanese, Korean, German, French, Spanish, Portuguese, Arabic, Hindi, Russian

Quality guardrails

• Pydantic v2 validation on all agent outputs
• Citation integrity check, hallucinated source ID detection, scoring formula correction
• Auto-retry (up to 2×) on validation failure

LLM support

• DeepSeek V3 · OpenAI GPT-4o · Anthropic Claude — auto-detected from API key

Benchmark suite

• 10 preset topics across Biotech, CleanTech, Energy, Materials, Cybersecurity, FoodTech
• Automated TRL range check, formula verification, section completeness check

Getting started

uv sync
cp .env.example .env   # add DEEPSEEK_API_KEY / OPENAI_API_KEY / ANTHROPIC_API_KEY + SERPER_API_KEY
uv run python app.py   # open http://localhost:7860