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. Auditing reward hackability in code RL training environments. 28.5% hackability on SWE-bench Verified; +14.14pp Pass@1 inflation across 134 model submissions.
[arxiv]
Arithmetic Intensity Aware Quantization. Post-training quantization optimizing FLOPs/byte. 1.66x speedup.
[arxiv]
Uncertainty Gated Retrieval. Test-time retrieval for semantic segmentation. +11% accuracy, 87.5% cost reduction.
[arxiv]
Multiver. Multi-agent memory-augmented code verification. Beats fine-tuned models on real-world vulnerability detection.
[arxiv]
Multi-Agent Code Verification. Proved why multi-agent verification works. 76% bug detection in <200 ms.
[arxiv]
Social Contagion and Bank Runs. LLM agents predicting the SVB→First Republic cascade.
[arxiv]
Test-Time Selection for VLAs. A perfect world model barely beats best-of-1; the bottleneck is the policy, not model fidelity.
[writeup] [github]
Geodesics on Finite 2-Manifolds. Discrete analog of geodesic spreading in general relativity.
[paper]
Mixing Rates of Markov Chains. Spectral gap controls how fast chains forget, with an RL interpretation.
[paper]
Projects
LoopHealth. A self-consensus RL loop halves its coverage while every metric looks healthy; a per-prompt signal catches it.
[writeup] [github]
Harvy. AI academic advisor with daily active usage from over 50% of Harvard students.
[site] [press]
Visdep. Codebase visualizer that turns large repos into interactive maps.
[site] [writeup]
STAR. Self-learning system that cuts human labeling for code generation by 85%.
Sisyphus. Physics-integrated world models for robot RL environments.
Robomerge. First autonomous VLA preprocessing pipeline.
[github]
Writing