# KIMA Thierry Tewendé - **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon) - **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST) - **Where:** Online - **Team:** [Thierry Tewende Kima](https://cerebralvalley.ai/u/Tewende) - **GitHub:** https://github.com/ThierryTe/ai-policy-invetigator - **Demo video:** https://www.youtube.com/watch?v=5Wn6jKOish8 - **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery - **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/529 AI POLICY INVESTIGATOR When a policy changes, which procedures need attention — and what is the evidence? AI Policy Investigator is an AI-powered application that helps organizations identify inconsistencies between updated policies and the operational procedures that must follow them. THE PROBLEM Organizations regularly update internal policies to reflect new regulations, business requirements, and operational decisions. However, existing procedures may continue to reference outdated rules, creating compliance risks, inconsistent practices, and additional manual review work. Traditional document comparison tools can highlight textual differences, but they do not necessarily explain how a policy change affects related procedures. OUR SOLUTION AI Policy Investigator uses Meta Muse Spark 1.3 to analyze a baseline policy, its replacement, and up to four selected procedure versions. The application extracts structured requirements, identifies policy changes, and evaluates their potential impact on existing procedures. It helps reviewers answer three essential questions: What changed between the original policy and its replacement? Which procedure requirements may now be inconsistent or incomplete? What original document evidence supports each finding? KEY FEATURES Evidence-based analysis: Findings are linked to source quotations and original PDF documents. Policy change detection: Identifies changes in obligations, responsibilities, deadlines, exceptions, and effective dates. Procedure impact assessment: Classifies potential issues as contradictions, missing requirements, ambiguities, or no conflict. Interactive investigation graph: Visualizes relationships between policy versions, requirements, and affected procedures. Human-in-the-loop review: Analysts can confirm, dismiss, or annotate findings without altering the original AI output. Asynchronous processing: Background workers handle investigations while preserving progress and recoverable checkpoints. Document traceability: Immutable document versions preserve the evidence used in each investigation. HOW WE USE META MUSE SPARK Meta Muse Spark 1.3 is the core semantic reasoning engine. Through the Meta Model API, the backend uses Muse Spark to extract structured requirements, align changes across policy versions, and assess their impact on selected procedures. The application combines model reasoning with strict backend validation of structured outputs, document provenance, source quotations, dates, and unknown fields. Invalid responses are rejected rather than presented as verified findings. Spring AI 2.0.1 integrates local embedding models through its EmbeddingModel abstraction. Local ONNX embeddings and PostgreSQL with pgvector support document retrieval independently of Meta's reasoning capabilities. TECHNICAL ARCHITECTURE Frontend: Angular 21 Backend: Java Spring Boot AI integration: Spring AI 2.0.1 for local embeddings AI reasoning: Meta Muse Spark 1.3 via Meta Model API Embeddings: Spring AI Transformers, local ONNX, all-MiniLM-L6-v2 Database: PostgreSQL with pgvector Processing: Asynchronous investigation workers Deployment: Docker Compose, GitHub Actions, and GitHub Container Registry LIVE DEMO — JUDGE ACCESS Application: https://ai-policy.tewendelabs.com Demo video: https://youtu.be/5Wn6jKOish8 Demo email: judge-test@tewendelabs.com Demo password: HackathonM@2026! INVESTIGATION 1 — REMOTE WORK POLICY (RECOMMENDED) Investigation ID: d706af38-a477-4177-bbb8-07001b0041bb This investigation compares a fictional 2025 remote-work policy with its 2026 replacement and examines the impact on an existing management procedure. The updated policy reduces remote-work allowances from three days to two, introduces mandatory office attendance on Tuesdays and Thursdays, and transfers exception approval from managers to Human Resources. This is the scenario featured in the demo video. INVESTIGATION 2 — EMPLOYEE CONTRACT Investigation ID: 44b5b0bc-920d-4e3f-8588-848c5b78e9e2 An additional document investigation demonstrating the application's evidence-based review workflow. RECOMMENDED EVALUATION 1. Sign in to the live application using the demo credentials. 2. Open the Remote Work investigation using the provided ID or its corresponding investigation link. 3. Explore the policy changes, procedure findings, original PDF evidence, and investigation graph. 4. Optionally review the Employee Contract investigation. 5. Starting a new investigation is optional and subject to available Meta API usage limits. Previously completed investigations can be reviewed without triggering new Meta inference calls. SOURCE CODE Backend: https://github.com/ThierryTe/ai-policy-invetigator Frontend: https://github.com/ThierryTe/ai-policy-invetigator-frontend IMPACT AND VISION AI Policy Investigator demonstrates how Meta Muse Spark can support a practical, evidence-based approach to policy governance. Instead of replacing compliance professionals, it helps them focus on the changes that matter, understand their operational consequences, and make informed decisions using traceable evidence. Our goal is to make policy reviews faster, clearer, and more actionable — while keeping humans responsible for the final decision. ## More from Meta Global AI Developer Hackathon - [HackX](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/526) - [Rui Codex](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/527) - [Vigiax](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/528) - [ExtBrain](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/530) - [Adrian](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/531) - [Centinela](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/532) --- Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/529. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt