I build systems that let AI agents do real work without anyone losing track of what they did, what it cost, or who approved it.

Right now that means three things: AgentOS, a governed control plane where every agent is budgeted, policy-bound and audited by construction; AlgoDesk, an options research and execution desk for the Indian markets where the money path stays deterministic and the models only ever propose; and Snipora, which turns a day of AI-industry video into an hour of reading.

I write here about what actually happened while building them — the failure, the mechanism underneath it, and what changed as a result. Mostly the parts I got wrong first.