Shreshth Rajan
I study Pure Math and Philosophy at Harvard. I took leave to be the first person on the team at Forge and founding engineer at Pillar.
Research
Envaudit. Audited code RL environments; 28.5% of SWE-bench Verified is reward-hackable.
[arxiv]
Arithmetic Intensity Aware Quantization. Quantized by FLOPs per byte for a 1.66x speedup.
[arxiv]
Uncertainty Gated Retrieval. Retrieved only when unsure: +11% accuracy, 87% cheaper.
[arxiv]
Multiver. Gave verification agents shared memory and beat fine-tuned detectors.
[arxiv]
Multi-Agent Code Verification. Proved why weak checkers beat strong ones; 76% of bugs in 200 ms.
[arxiv]
Social Contagion and Bank Runs. Modeled the SVB to First Republic cascade with LLM agents.
[arxiv]
Test-Time Selection for VLAs. Showed a perfect world model barely beats one sample.
[writeup] [github]
Geodesics on Finite 2-Manifolds. Built a discrete analog of geodesic spreading in relativity.
[paper]
Mixing Rates of Markov Chains. Bounded how fast a chain forgets where it started.
[paper]
Projects
LoopHealth. Caught an RL loop halving its coverage while every metric stayed flat.
[writeup] [github]
Harvy. Built an AI academic advisor that over half of Harvard uses daily.
[site] [press]
Visdep. Turned large repos into dependency maps you can navigate.
[site] [writeup]
STAR. Cut human labeling for code generation by 85% with a self-training loop.
Sisyphus. Built physics-integrated world models for robot RL environments.
Robomerge. Built an autonomous preprocessing pipeline for robot training data.
[github]
Writing