The Story#

How a Database Systems assignment became irag — by Karan Garg.

I had a Database Systems assignment coming up and honestly had no idea what to build. Around then, Andrej Karpathy put out his LLM Wiki, so I started tinkering with it and found the issues fast: it was all just files — nothing versioned or checked, nothing stopping it from quietly going stale while people kept trusting it. I figured I'd try fixing that for the assignment.

I picked a law firm as the domain because it sounded fun to build, and it turned out to be the right kind of hard too. Law firms don't forgive mistakes — a wrong court date or a missed conflict of interest is a real failure, not a bad chatbot answer. Building under that pressure got me to the actual idea:

It should be a database system with an AI layer, not an AI system with a database attached.

Five tables did the real work underneath — pages, revisions, links, events, contradictions — while the database handled the bookkeeping and the AI only ever wrote prose. I tested it by imagining the AI ripped out completely: a full working law-firm system was still standing there.

That's when it clicked that those five tables were never about law firms at all. They were an answer to something bigger: how do you let an AI write things you can actually trust? So I stripped the law-firm parts out and pointed the same core at a problem that had been bugging me the whole time — coding agents that forget everything the second you close the terminal.

Around then I noticed tools like Graphify and claude-mem popping up too — other people circling the same problem from different angles. It felt like confirmation I was onto something real.

That became irag. One SQLite file in your repo — no service, no cloud, no keys. The agent stops grepping around; its map, search, and context are SQL now, not more model calls. Memory that can't quietly lie — every claim gets checked against the real code and gated in CI. Reads cost next to nothing: a few thousand tokens of briefing instead of re-exploring everything. A new chat resumes an old project in about a hundred and fifty tokens, like no time passed.

Where to go from here: