Gourav

Airlock — Secure Invoice Intelligence and Payment Control Meta Global AI Developer Hackathon 2026 Track: Indie Theme: Build for Your World An invoice can request a payment. It cannot authorize one. OVERVIEW AND IMPACT Airlock is a multimodal invoice-verification and payment-control application for business owners, bookkeepers, and small finance teams. It addresses bank-account substitution, duplicate billing, purchase-order discrepancies, inconsistent totals, and prompt injection. AI interprets evidence, deterministic software checks financial claims, and the owner authorizes the exact proposal. Execution is mock-only; no funds move. RESEARCH FOUNDATION Meta’s Agents Rule of Two informs Airlock’s separation of untrusted input from financial authority. The supplied research cites approximately $3.05 billion in 2025 FBI IC3 business-email-compromise losses—not AI-injection losses or savings achieved by Airlock. https://ai.meta.com/blog/practical-ai-agent-security/ https://www.ic3.gov/AnnualReport/Reports/2025_IC3Report.pdf META PRODUCTS Muse Spark 1.3 (muse-spark-1.3) extracts structured invoice candidates. Muse Voice Transcribe (muse-voice-transcribe-1.0) supplies transcripts for conflict checks. SAM 3.1 (sam-3.1) supplies delivery geometry and count observations, never proof of authenticity. Muse Spark Open is integrated through Baseten; Glimmer has an optional local adapter, with no verified local inference run. Muse Code supported implementation and debugging. Successful live API smoke checks were reported for Spark, SAM, and Voice; these establish connectivity, not end-to-end accuracy. TECHNICAL EXECUTION Python, FastAPI, Pydantic v2, React 19, TypeScript, SQLite, and httpx connect validated facts to trusted vendor, bank, and purchase-order records. Controls include Decimal reconciliation, Unicode normalization, Levenshtein ambiguity detection, HMAC-SHA-256 bank fingerprints, duplicate keys, Argon2 PIN hashing, proposal-bound single-use challenges, expiry checks, idempotency, and hash-chained audit events. EDGE CASES AND EVIDENCE Rules R1–R10 cover identity, bank mismatch, arithmetic, duplicates, limits, dates, quantity conflicts, missing fields, verifier errors, and ambiguity. Provider failures cannot silently become approval. Reported synthetic evaluation: 19/19 expected outcomes, zero false-ALLOW across 10 adversarial cases, 15/15 schema checks, and all ten rules exercised. A later supplied audit reports 950 backend and 61 frontend tests passed. These are local results, not live accuracy guarantees. Auth gaps, CI failures, and licensing require remediation before production use. Proposed growth: B2B subscriptions, controlled pilots, and accounting integrations. GitHub: https://github.com/Gouravjaat07/Airlock-Invoice-Security-Meta-AI-Hackathon-2026