Compliance intelligence stack

EHS Intel Platform

DB-first environmental, health, and safety intelligence that connects public agency data to retrieval tools and a streaming AI interface.

Back to main page

Product Surface

EHS Intel platform product preview
Static project preview for the compliance intelligence interface.
EHS Intel architecture diagram
Agency datasets, PostgreSQL, MCP/FastAPI tools, and streaming Next.js AI interface.

Problem

EHS research often requires stitching together OSHA, EPA, ITA, TRI, MSHA, Superfund, and other public records before a company or facility risk picture becomes usable. A plain chatbot is not enough because compliance answers need source-backed retrieval and careful claim boundaries.

Solution

The documented architecture is a PostgreSQL-centered intelligence platform with ingestion and sync scripts, normalized entity records, indexed query paths, source-fidelity fields, an MCP/FastAPI tool server, and a Next.js interface that streams model responses backed by resolve and brief endpoints.

Retrieval Flow

EHS Intel retrieval flow diagram
Company queries resolve entities first, retrieve supporting agency records, then synthesize guarded brief output.

Design Decisions

DB-First Answers

Compliance-facing responses are grounded in indexed database retrieval before any synthesis layer is allowed to summarize.

Inspectable Resolution

Entity matching is separated from brief generation so company resolution can be reviewed before conclusions are drawn.

Source Fidelity

Raw-row style fields and provenance details preserve the path back to source agency records.

Guarded Synthesis

Prompt constraints are designed to avoid invented violations, unsupported red flags, or regulatory claims beyond retrieved records.

Operations And Reliability

The documented local workflow starts PostgreSQL in Docker, runs ingestion and migration paths, launches the MCP/FastAPI HTTP server, and serves a Next.js/React/TypeScript frontend. The web route coordinates streaming tool calls against retrieval endpoints rather than relying on free-form model context alone.