Company
Pramaana Labs is a frontier AI lab building the verification layer for AI. Founded in 2025 and headquartered in Palo Alto, the company turns complex human knowledge, including tax codes, clinical guidelines, legal rules, and safety constraints, into machine-checkable logic, so every answer can be traced, challenged, and proved. The company was founded by a team out of Google, Google DeepMind, and Glean, and is backed by Khosla Ventures, with participation from Accel, Nexus Venture Partners, Premji Invest, and others as well as angels such as Pushmeet Kohli and Sriram Rajamani.
Pramaana's architecture pairs foundation models trained to formalize and reason with a symbolic world model encoded in the Lean proof language. Regulatory, statutory, policy, and scientific text is converted into formal representations, and the system returns proof artifacts that domain experts can inspect. Accountability is built into the architecture rather than added afterward through policy or prompting.
The role
You'll be a Member of Technical Staff on the AI team - the group responsible for the models underneath Panini and Hardy. This is a hands-on research role: you'll own experiments end-to-end, from the idea through the training run, the debugging, the eval, and the honest write-up of what did and didn't work.
What you'll actually work on
Concretely, some mix of:
-
Post-training and RL for proof search. Reward design over a verifier that gives you a real, non-gameable signal - one of the few places in ML where ground truth is actually ground truth.
-
Distillation and capability transfer. Closing the gap between a large teacher and a model small enough to run a search loop economically.
-
Autoformalization. Getting models to translate natural-language statements - a statute, a spec, a theorem - into formal representations that typecheck and mean what the source meant.
-
Inference-time search. The compute budget for a proof attempt is a design parameter. Spending it well is a research problem.
-
The infrastructure that makes the above real. Training loops, data pipelines, evals you'd actually trust.
What we're looking for
We hire for depth over pedigree. We care about what you've actually built and understood, judged on the work itself - not where you did it or how many things you've touched.
-
Technical depth. You understand your own work from first principles. You know why you made the choices you made, not just what you did and that holds up under questioning rather than getting vague.
-
You build and ship. You can implement, train, debug, and scale your own ideas end-to-end, including the infra reality of pre-training and RL loops. You don't need someone else to make the idea real.
-
Rigor. You care instinctively about sound claims, verification, and reproducibility. We come from a world where correctness is the whole game; overclaiming is disqualifying here in a way it isn't everywhere.
-
Agency. You drive toward outcomes and follow through. You don't wait to be told the next step, and you don't let things quietly stall.
-
Perseverance. You've gone deep on something over a long stretch — and stayed with it after the first three ideas failed.
-
Alignment. You want to work on this problem. We'd rather hire someone convinced this is the most important thing they could be doing than someone excellent who's shopping.
-
No jerks. Someone who makes the people around them worse is a net negative no matter how strong they are individually. Talent doesn't buy a pass.