Track 01 — Data Rescue | M-AGENTS Hackathon | June 7, 2026

🎬 Demo video: https://app.trupeer.ai/view/5U2nTdaab/asset-guard-ai-user-manual

A 4-agent system that rescues corrupted manufacturing data and produces a signed-ready, plain-English audit narrative — operable by a compliance officer who has never opened a database.

cd backend
pip install -r requirements.txt
cp .env.example .env        # Windows: copy .env.example .env

AssedGuard runs fully offline with no keys — all detection, ranking, and fixing is deterministic Python, and the narrative falls back to a deterministic template. Add keys only to enable Claude's narrative polish and Cognee cloud:

• Anthropic: https://console.anthropic.com
• Cognee (free 14-day trial): https://www.cognee.ai

Put them in backend/.env:

ANTHROPIC_API_KEY=sk-ant-...
COGNEE_API_KEY=...

cd backend
uvicorn main:app --reload --port 8000

Open frontend/index.html directly in your browser (no build step — CDN React).

• Drag data/sample.csv into the upload zone
• The audit runs automatically — watch the 4 agents fire in sequence
• Review the ranked findings table (click any row to expand the reason)
• Click Download Audit Narrative PDF

CSV ─► Scout ─► Ranker ─► Fixer ─► Narrator ─► PDF
        │         │         │          │
        └─────── Cognee shared memory ─┘
        (each agent writes its output; the next reads it before acting)

• Agent 1 — Scout detects 6 issue classes with deterministic Python: exact duplicates, conflicting lot quantities, unit conflicts, statistical outliers (>3σ), missing timestamps, and compliance contradictions (PASS + out-of-spec temperature). HIGH findings are enriched with the specific regulation violated. Out-of-range sensor readings are additionally scored by a robust Bayesian model (PyMC) — a Student-T fit that resists being skewed by the outliers themselves, giving each reading a calibrated anomaly probability and the sensor's 95% credible normal range. It is best-effort: if PyMC is unavailable the deterministic 3σ result stands, so detection never breaks.
• Agent 2 — Ranker orders findings by a transparent rubric (compliance contradictions → traceability → measurement). Every rank has a logged plain-English reason. Claude peer-reviews the ordering and attaches a note.
• Agent 3 — Fixer auto-fixes (dedupe keep-most-recent, flag missing dates), flags (unit conflicts, outliers), and escalates (contradictions, lot conflicts) — each with a logged reason — and emits a corrected CSV.
• Agent 4 — Narrator reads the full pipeline memory and writes a 6-section audit narrative, rendered to a professional PDF with reportlab.

Memory layer: Cognee is the handoff bus — every agent writes its structured output and the next agent reads it before acting (see memory/cognee_store.py: write_memory calls cognee.add() when a COGNEE_API_KEY is configured). By design, if the key/SDK is unavailable the same read/write API transparently mirrors into an in-process store, so the handoff contract is identical and the pipeline never breaks during a live demo. To run against the real Cognee cloud, pip install cognee and set COGNEE_API_KEY in backend/.env.

• Cognee — shared memory layer between all 4 agents (real read/write handoffs)
• Trupeer — demo video (see submission)
• Geodo — regulatory entity research by the Domain Expert

1. Multi-agent system — 4 specialized agents with real Cognee handoffs; each reads upstream memory before acting and writes its output after.
2. Solves a real data-rescue problem — duplicates, unit conflicts, outliers, missing data, and compliance contradictions in factory data.
3. Usable by a non-technical end user — drag a CSV, one automatic run, download a PDF. Zero jargon, zero engineering required.
4. Tangible deliverable — a signed-ready audit narrative PDF plus a corrected dataset CSV.
5. Every decision is explainable — every ranking and fixing action carries a visible, deterministic, plain-English reason. No decision is ever an unexplained model verdict.