We build Sarathi, a local-first coding agent that plans, writes, and verifies work against real runtimes on your machine.
Everything we ship is downstream of these. They are not marketing — they are the filter on every design decision.
We begin with first principles, then test them against evidence. Mathematical elegance is a hypothesis until independent results show that it works.
Software that runs on your machine, against your keys, on your terms. No silent telemetry, no rented intelligence. Your work never leaves your custody unless you choose.
A small lab shipping tools we use ourselves. No engagement metrics, no growth loops. We measure success in decades of useful software, not quarters of attention.
Useful software should respect your machine, survive real verification, and make difficult work meaningfully easier.
Our research and engineering share one discipline: claims must survive contact with ground truth.
In code, the runtime is the test. In research, independent evidence is the test. We keep the claim smaller than the proof.
Treating code output as adversarial: launch real runtimes, probe endpoints, fail loudly. Verification is not a second model — it is ground truth.
Agent loops that plan, execute, and recover on-device. Long-horizon memory and crash recovery that resumes exactly where the queue stopped.
We take on a small number of research and engineering engagements where the problem is hard and the math matters. Reach out.
lab@bhaskara.io