Hackathon Portal
AI Tinkerers - Boston
Team

Governed Incident Response

Project Concept

Governed Incident Response is an AI agent for workplace safety incidents in regulated industries. The agent handles emergency procedure lookup, severity-1 notifications, and procedure update drafting. Every action runs through a governance layer before it executes: role-based access control, retrieval scoring, evidence gating, hallucination detection (HHEM), and an immutable audit chain. The right-side instrument panel shows the governance cascade in real time. The result is an AI agent that helps at the speed of an emergency and stays auditable at the standard of regulation.

Demo: three queries, four governance outcomes

SERVE — Operator asks about atmospheric testing for confined space entry. RBAC passes, retrieval finds the procedure, ACL clears it, evidence gate passes, HHEM passes, audit chains. Cited response delivered.

SERVE + BLOCK + ROUTE — Operator queries a confined space collapse at a petroleum facility. Two procedures retrieved and served. The agent attempts a severity-1 notification. Operator role is not authorized. BLOCKED. The notification routes to supervisor approval queue. Three governance outcomes in one query.

REFUSE — Operator asks about TIER greenhouse gas reporting. Evidence gate scores 0.41 against a 0.70 threshold. The system refuses rather than hallucinate.

A replay fallback runs offline against hardcoded event sequences for venue WiFi failures.

Stack

CopilotKit 1.56.5, Next.js 16, TypeScript, Tailwind, Zustand, OpenAI GPT-4o.

Solo build by Arnaldo Sepulveda.

Entry

Status: Submitted

Last saved: May 09 at 5:34 PM EDT

Team Roster

Message board not available for this team yet.

Arnaldo Sepulveda Team Lead RSVP Approved

Lead Engineer at Keystone Applied Intelligence
Arnaldo Sepulveda (Lead): Solo build. End-to-end ownership of all components. Frontend & Generative UI: - Two-column 60/40 layout (CopilotChat left, ControlFeedback right) - Three CopilotKit useCopilotAction tools: lookup_procedure, queue_notification, draft_procedure_update - Custom render callbacks rendering color-coded procedure cards (approved/denied/fail-closed) - ControlFeedback instrument panel with pulsing PENDING-to-PASS HHEM animation - Replay fallback mode with hardcoded event sequences for offline demo Governance layer: - Role-based access control (operator, supervisor, admin) with isAuthorized logic - Evidence threshold gate, ACL filter, HHEM hallucination scoring - Immutable hash-chained audit trail (in-memory) - Four-outcome controller decision space: SERVE, REFUSE, BLOCK, ROUTE State & integration: - Zustand event store with addEvent (returns ID) / updateEvent pattern - CopilotKit OpenAI adapter (gpt-4o) - Strict CopilotKit 1.56.5 pinning (no upgrade during hackathon)
Enterprise AI engineer with 12+ years at Genesys building production systems for regulated and public sector customers. Path ran from SIP infrastructure through cloud migrations through leading a support engineering team for enterprise Cloud CX. MScE in Electrical Engineering with published research in AI/ML for smart grid optimization. Most recently built a governed retrieval and agent system that runs entirely on-premises with no external API dependencies. It enforces role-based access control at query time, ties answers to source evidence, refuses when confidence is insufficient, extends the same governance model to tool-using agents, and produces tamper-evident audit trails. Connecting with engineers working on enterprise AI, knowledge systems, and retrieval infrastructure.
Evaluation methodology for RAG systems beyond basic metrics. How teams measure retrieval quality degradation in production. Fail-closed and abstention design patterns for high-stakes AI. Access control architectures for multi-tenant knowledge systems. On-premises LLM deployment at scale. How enterprise AI teams handle document versioning and point-in-time retrieval. Connecting with engineers at companies building AI for healthcare, legal, defence, energy, and compliance.
Built a governed retrieval and agent system for regulated industries. Retrieval: 11-step pipeline, hybrid retrieval (pgvector + FTS), query-time RBAC, evidence thresholding, fail-closed gate, hash-chained audit. One production deployment. Agent extension: per-step authorization, HITL routing, evidence gating, evaluated 186 cases / 0 fail. Stack: Python, FastAPI, PostgreSQL + pgvector, Ollama, React, Docker. Eval: P@1=0.75, MRR=0.79, 0 ACL leaks.