Fraudshield Loop

# /fraudshield-loop
Run for exactly 500 iterations. Stop early ONLY if user says "stop" or "exit".

Context: FraudShield — AI Fraud Detection API & Dashboard
Stack: Python · XGBoost · SHAP · FastAPI · React · Docker
Target: Win 5 concurrent hackathons. Judges: Goldman Sachs, Citibank, ByteDance.

---

**Before every iteration:**
```bash
COUNT=$(cat ./loop-logs/.iteration-count 2>/dev/null || echo 0); N=$((COUNT + 1))
echo "Iteration $N / 500"

Read ALL ./loop-logs/summary-*.md — recall full history, avoid repeating failures. Check DEPLOY_URL.txt. Probe live API if it exists.

---

STEP 1 — Autonomous Analysis + Plan

You are a senior ML engineer AND a Goldman Sachs quant judging this submission. Do NOT blindly follow the PRD. Think independently.

Ask yourself:

• What is the SINGLE biggest gap between current state and winning?
• Would a Citibank judge be impressed by the SHAP output right now?
• Is the < 10ms latency claim actually benchmarked?
• Does the dashboard look like a real fintech tool or a student project?
• Is anything broken that makes a judge close the tab immediately?
• What is NOT in the PRD that would dramatically improve the score?
• What did previous iterations fail at — how do we not repeat that?

Use sequential-thinking MCP for complex ML/architecture decisions. Use context7 MCP to verify FastAPI, XGBoost, SHAP, React APIs before writing.

Write ITERATION PLAN → ./loop-logs/plan-$N.md:

• Current state (what works, broken, missing)
• One highest-leverage improvement
• Exact files to change
• How you will verify success

PRD is a reference, not a ceiling. If something better emerges, build it and update PRD.

STEP 2 — Self-Verify

Review as the strictest judge. FAIL if:

• Vague plan with no file paths
• Ignores critical bug in favor of polish
• Introduces SMOTE, serves model.pkl via API, re-fits scaler at inference
• Claims < 10ms without a benchmark plan
• Repeats a previous iteration exactly
• Unmeasurable success criteria

If FAIL: rewrite plan. Repeat until PASS.

STEP 3 — Build

Execute completely. No half-measures.

Delegate via subagents:

• ml/ changes → ml-engineer agent
• dashboard/ changes → frontend-agent agent
• After API/CORS/DB changes → security-reviewer agent

Autonomous rules:

• Fix bugs found along the way
• If better idea emerges mid-build, adapt and log why
• If PRD spec is wrong in practice, update both code and PRD

Standards:

• Python: type hints, no bare except, docstrings
• FastAPI: Pydantic on ALL inputs/outputs
• React: TypeScript strict, no any, all data-testid attributes present
• SHAP: TreeExplainer at module level — never per-request

git add -A && git commit -m "feat: iter $N — <what changed>"

STEP 4 — Test + Benchmark

Fix every failure before continuing.

# Type check
mypy api/ ml/ --ignore-missing-imports 2>&1 | tee ./loop-logs/mypy-$N.txt || true

# Unit + integration
pytest tests/ -v 2>&1 | tee ./loop-logs/pytest-$N.txt

# Latency — /latency-check skill
# Run /latency-check and save output to loop-logs/latency-$N.txt

# ML model regression (if model.pkl exists)
python3 -c "
import pickle, numpy as np, os
if not os.path.exists('ml/model.pkl'): print('model.pkl not built yet'); exit(0)
bundle = pickle.load(open('ml/model.pkl','rb'))
model, scaler = bundle['model'], bundle['scaler']
assert 0 <= model.predict_proba([[0]*29])[0][1] <= 1
fixed = np.array([[1200.0,-4.5,-3.0]+[0.0]*26])
score = model.predict_proba(np.hstack([scaler.transform(fixed[:,:1]), fixed[:,1:]]))[0][1]
print(f'Suspicious: {score:.4f}')
assert score > 0.5, f'Score {score:.4f} too low — check model'
print('MODEL PASS')
" 2>&1 | tee ./loop-logs/model-$N.txt

# Playwright dashboard
npx playwright test tests/dashboard/ --reporter=list 2>&1 \
  | tee ./loop-logs/playwright-$N.txt || echo "⚠ Dashboard tests failed"

Save screenshots → ./screenshots/

STEP 5 — Quality Review + Status + Deploy

API review:

curl -s http://localhost:8000/docs | grep -q "FraudShield" && echo "Swagger OK"
curl -s -X POST http://localhost:8000/predict \
  -H "Content-Type: application/json" \
  -d '{"amount":1200,"features":[-4.5,-3.0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0]}' \
  | python3 -m json.tool

Dashboard review: Look at screenshots. Check:

• Fraud rows: unmistakably red WITH secondary icon (not color alone)
• SHAP chart: readable to a non-ML person
• Overall feel: real fintech tool, not student project Fix anything that fails BEFORE writing status.

Write ./loop-logs/status-$N.md:

• What was built, what changed
• pytest / latency / Playwright results
• Model metrics if retrained
• Honest: "right now this would rank ___"
• Top 3 for next iteration

Run /submission-check skill — paste results in status log.

Deploy:

npm run build --prefix dashboard 2>/dev/null \
  && npx vercel --prod --yes --cwd dashboard 2>/dev/null \
  && echo "Deployed" >> DEPLOY_URL.txt \
  || echo "⚠ Deploy failed"

STEP 6 — Compact + Counter + Loop

Write ./loop-logs/summary-$N.md:

• Iter $N of 500
• Exact state: endpoints, test pass rates, current F1/latency
• What was built and whether it worked
• What FAILED and WHY (prevent repetition)
• Completion % toward Definition of Done
• Single most important thing for N+1

Update counter:

echo $N > ./loop-logs/.iteration-count

MANDATORY /compact — every single iteration, no exceptions. Context freshness is critical across 500 iterations.

Loop check:

[ "$N" -ge 500 ] \
  && echo "✅ 500 iterations complete." && cat ./loop-logs/summary-500.md \
  || echo "Restarting → iteration $((N+1))"

If N < 500: restart at STEP 1 immediately. Stop early ONLY if user says "stop" or "exit". ```