03 - Process Safety - Hybrid RAG
A process-safety workspace that turns HAZOP and LOPA methodology into structured, auditable software, with a grounded AI layer that suggests causes, consequences and safeguards from plant documentation.
HAZOP and LOPA work depends on structure: nodes, guidewords, deviations, causes, consequences, safeguards, risk ranks and approvals. When that structure lives in spreadsheets and long documents, it becomes hard to search, compare, reuse and audit.
PreventA keeps the methodology intact while moving the work into a product surface: each study is decomposed into typed records, risk state is visible, and AI suggestions stay clearly separated from approved engineering decisions.
I designed the product around the actual safety-study objects: studies, nodes, deviations, safeguards, risk matrix cells and source-backed recommendations.
The strongest screens are the HAZOP worksheet and risk matrix. Together they show the core product promise: keep the safety method structured while using AI to accelerate evidence gathering and first-pass recommendation drafting.
HAZOP workspace
The worksheet captures guideword, deviation, causes, consequences, safeguards and recommendation status in one review surface.
Risk matrix
The 5x5 matrix gives engineers a shared language for ranking severity and likelihood before and after safeguards.
Source grounding
The retrieval layer uses full-text and vector search, then keeps source-backed suggestions distinct from approved study rows.
The system is built so deterministic study data and probabilistic AI suggestions never collapse into one ambiguous blob.
Nodes, deviations, safeguards and risk cells are stored as structured records that can be filtered, audited and compared across studies.
Postgres full-text search and pgvector retrieval are fused to find relevant plant docs, incident notes and prior safeguards.
AI output enters as a suggestion with source context; human approval is required before it becomes part of the official worksheet.