From CAN Bus to RAG Pipelines: Why I Build

I didn't plan on ending up in GenAI. My plan, if you'd asked me a year and a half ago, was to be very good at making cars talk to each other correctly.

I joined Tata Technologies as an GET and became Associate Software Engineer working on AUTOSAR Basic Software configuration and ECU integration — CanIf, PduR, Com, DEM, DCM, the whole diagnostic stack. I spent my days in ISOLAR-AB and CANoe, writing CAPL scripts, chasing down UDS diagnostic sessions on a Raspberry Pi 4 I'd rigged up as a portable diagnostic tool. It's precise, unglamorous, high-stakes work — the kind where a misconfigured signal mapping can mean a safety function doesn't trigger when it needs to.

That background is exactly why I ended up building in GenAI, not despite it.

The problem that pulled me in

Functional safety documentation in automotive is enormous — standards, requirement traces, hazard analyses, compliance evidence, all scattered across formats and systems that don't talk to each other. I kept thinking: this is a retrieval problem. Not a "let's bolt on a chatbot" problem — a genuine, structured, audit-grade retrieval problem, in a domain where a wrong or hallucinated answer isn't a minor inconvenience, it's a compliance risk.

So I built FUSA: a multi-tenant, on-premises RAG pipeline for automotive functional safety compliance. It uses LangGraph for orchestration, Qdrant for hybrid retrieval — dense embeddings plus BM25, combined through Reciprocal Rank Fusion — a Kuzu graph database for relationship-aware context, and PostgreSQL underneath it all. On-prem wasn't a preference, it was a requirement: this data doesn't leave the building. Getting hybrid retrieval to actually outperform naive dense search, and getting multi-tenancy right without leaking context across teams, taught me more about production AI systems than any course could have.

That project became the bridge. Once I'd built one real thing, I wanted to build more.

Learning by shipping, not just studying

I went on to build an agent that connects to Atlassian's official remote MCP server — OAuth 2.1, LangGraph, live tool-calling against Jira and Confluence. It sounds small written down like that, but getting an agent to reliably read, reason over, and act on real enterprise APIs — not toy endpoints — surfaced problems no tutorial prepares you for: library version mismatches, transport limitations, the gap between "works in a notebook" and "works against a live system."

I also started an offline, multilingual RAG assistant for ASHA health workers in rural India — grounded in real Government of India community health guidelines, designed to work without reliable connectivity. It's early, but it's the project closest to why I care about any of this: the best AI systems are the ones that work for people who are usually an afterthought in how this technology gets designed.

What connects the two halves of my work

People sometimes ask why I haven't dropped the embedded/AUTOSAR side to go "all in" on AI. Honestly, I think it's my advantage, not a distraction. Safety-critical systems engineering teaches you to distrust your own system by default — to ask "what happens when this is wrong," not just "does this work on the happy path." I bring that instinct into how I build with LLMs: I'd rather have a RAG pipeline that says "I don't have enough grounding to answer that" than one that sounds confident and is wrong.

Why I'm here

I'm mostly self-taught in GenAI — no bootcamp, no formal ML degree, just a functional safety problem that needed solving and a lot of stubbornness. Cohort 3 is the first time I'll be building alongside people who've been doing this deliberately, at scale, with infrastructure and mentorship I don't have access to on my own. I want to get better at building agentic systems that are genuinely trustworthy in domains where trust isn't optional — safety, healthcare, compliance — and I want to stop learning that in isolation.

If any of this resonates — if you're also somewhere between two disciplines, using one to make the other more rigorous — I'd love to compare notes.

Rithika G
Associate Software Engineer, Tata Technologies | Building FUSA, an MCP-powered Jira agent, and a health RAG assistant for rural ASHA workers*