1# HARA AI
2
3Safety-relevant AI system for Hazard Analysis and Risk Assessment
4
5## Overview
6
7HARA AI automates the identification of safety goals and ASIL (Automotive
8Safety Integrity Level) classifications per ISO 26262. It replaces weeks
9of manual hazard analysis with an AI pipeline that reasons over vehicle
10functions, operational situations, and failure modes.
11
12## How it works
13
141. Parse vehicle function definitions and operational design domains
152. Generate hazardous events using LLM reasoning chains
163. Classify severity, exposure, and controllability per ISO 26262 tables
174. Output ASIL ratings (A-D / QM) with full traceability
18
19## Tech Stack
20
21| Layer         | Tech                            |
22| ------------- | ------------------------------- |
23| AI Engine     | LangChain, RAG pipeline         |
24| LLM           | Fine-tuned Qwen 7B (QLoRA)      |
25| Backend       | FastAPI, Python 3.11             |
26| Safety Std    | ISO 26262 Part 3 (HARA)         |
27| Evaluation    | LangSmith, custom ASIL metrics   |
28
29## Key Results
30
31- 90%+ accuracy on ASIL classification vs. human expert baseline
32- Reduced HARA cycle time from 2 weeks to 2 hours
33- Full audit trail for safety certification workflows
34
35## Architecture
36
37```
38Vehicle Functions → Hazard Identification → Risk Assessment → ASIL Rating
39       ↓                    ↓                     ↓              ↓
40  ODD Parser         LLM Chain (CoT)        ISO 26262       Report Gen
41                     + RAG Context           Tables          + Tracing
42```
43
44
45Built at Bosch Global Software Technologies | Lead Engineer - Gen AI