The PRD review tool for product managers who don't have time to wait for a senior PM to read their doc.

Paste a PRD. Get a structured 100-point score across 10 dimensions, the single biggest gap, three specific rewrite suggestions, and the questions a reviewer would actually ask. Profile-aware: a startup PRD, a TCS engagement doc, and a Big Tech design doc are scored differently.

This is the production product. The repository also contains experimental modules (task breakdown, assignment, sprint tracking) — see Roadmap for current status.

Two modes for the analyzer:

Profile-aware scoring for:

• startup — falsifiability, lean PRD conventions
• it_services — SOW alignment, client-side acceptance, KT plan
• big_tech — design-doc conventions, alternatives mandatory, OKRs
• financial_services — regulatory considerations (see Compliance Caveat)

Three modes, in order of friction. Pick whichever fits.

pip install -r requirements.txt
python -m eval.runner --mode rules_only --max 3   # see it working in 1 second

Score any PRD with the deterministic regex + heuristic engine (82% band-pass on our 50-fixture eval — see LEADERBOARD.md). Free, offline, private. Useful as a CI gate or a fast pre-check.

from engine.module3_analyzer.prd_analyzer import PRDAnalyzer
analyzer = PRDAnalyzer()                       # no LLM client
result = analyzer.analyze(open("my_prd.md").read(), mode="rules_only")
print(f"{result['total_score']}/100 — {result['rating']}")

# Install Ollama from https://ollama.com, then:
ollama pull llama3.1:8b
export AEGIS_LOCAL_BASE_URL=http://localhost:11434/v1
export AEGIS_LOCAL_MODEL=llama3.1:8b
python -m engine.verify --live                # confirms local model responds

Same JSON output shape as the cloud path, but runs entirely on your hardware. Slower (~30-60s per PRD on CPU) but unlimited and fully private.

Pick either of these free providers — no credit card:

# Option A: Groq (recommended — fast, generous daily limit)
echo "GROQ_API_KEY=gsk_xxxxxxxxxxxx" >> .env       # get one at console.groq.com

# Option B: OpenRouter (free models like deepseek-chat-v3:free)
echo "OPENROUTER_API_KEY=sk-or-xxxxxxxxxxxx" >> .env  # openrouter.ai

python -m engine.verify --live

Aegis auto-detects whichever key is present. Set AEGIS_REDACT=1 if you handle PII or regulated data — see SECURITY.md. For detailed signup steps and per-provider rate limits, see BUDGET.md.

from engine.ai_engine import AegisEngine

eng = AegisEngine()

# Free, instant: structural scan (no LLM)
scan = eng.scan_prd(open("my_prd.md").read())
print(f"Completeness: {scan['completeness']}%")
print(f"Missing: {scan['missing_sections']}")

# Full analysis (uses whatever's configured: rules, local, or cloud)
result = eng.analyze_prd(open("my_prd.md").read())
print(f"Score: {result['total_score']}/100 — {result['rating']}")
print(f"Critical gap: {result['critical_gap']}")
for tip in result['top_improvements']:
    print(f"  - {tip}")

       ┌──────────────────┐
       │  PRD text input  │
       └────────┬─────────┘
                │
     ┌──────────┴──────────┐
     │                     │
     ▼                     ▼
┌──────────┐         ┌──────────────┐
│  Layer 1 │         │  Layer 1+2+3 │
│  scan()  │         │   analyze()  │
│  regex   │         │   regex →    │
│  <100ms  │         │   LLM →      │
│  free    │         │   RAG ground │
└──────────┘         │   ~3s, $0.03 │
                     └──────┬───────┘
                            │
                            ▼
                     ┌──────────────┐
                     │ 100-pt score │
                     │ 10 dims      │
                     │ rewrites     │
                     │ Q&A          │
                     └──────────────┘

Underneath:

• Module 1 — RAG Knowledge Base. Hybrid BM25 + dense vector + Reciprocal Rank Fusion + cross-encoder re-ranking, over a 99K-word curated PM corpus.
• Module 3 — Analyzer. 10-dimension rubric (problem statement, target user, goals, metrics, solution, risks, stakeholders, open questions, launch, writing). Profile-specific scoring overlays.
• LLMClient. Provider-agnostic (OpenAI / Anthropic / Azure / local OpenAI-compatible). Built-in redaction, audit logging, and local-model fallback. See SECURITY.md.

Every prompt change is gated by an eval set:

# Structural eval (no API key, runs in CI)
pytest eval/test_eval.py -v

# Full LLM eval (10 anchors × 5 perturbations = 50 fixtures)
# Costs $0.00 if you use Groq / OpenRouter / Gemini free tier — see BUDGET.md
python -m eval.runner

See eval/README.md for acceptance thresholds.

Experimental modules are gated behind the env var AEGIS_ENABLE_EXPERIMENTAL=1. Do not deploy experimental modules to production.

The financial_services profile enriches scoring with finserv-specific heuristics — it does not make this system finserv-compliant.

In particular, the system as shipped:

• Sends PRD content to a third-party LLM unless redaction is enabled.
• Has no guaranteed data residency.
• Does not produce a regulator-grade audit trail.
• Has not been independently security-reviewed or SOC2-attested.

If you operate under SEC, FINRA, MiFID, GDPR-with-finserv-overlay, or similar regimes, do not use the LLM-backed analyze path on material non-public information without:

1. Setting AEGIS_REDACT=1 (see SECURITY.md).
2. Configuring a local model fallback (AEGIS_LOCAL_BASE_URL).
3. Independent legal sign-off on the data flow.

For a one-page customer-facing summary aimed at compliance officers, see TRUST.md. For the full technical detail (redaction patterns, audit-log schema, env vars, every guarantee + every non-guarantee), see SECURITY.md.

# Run all tests (no API key required for unit tests)
pytest tests/ eval/test_eval.py -v

# Eval the analyzer end-to-end
python -m eval.runner --scan-only       # CI mode, no API
python -m eval.runner                    # Full mode, costs ~$0.15 with gpt-4o-mini

When you change a prompt:

1. Run the full eval.
2. Commit the resulting eval/results.json.
3. Diff band_pass_rate and dim_* counters against the previous baseline.
4. Don't merge a regression without an explicit reason in the PR.