Selected builds
The public layer of the work: runnable artifacts, evaluations, and small systems that make a research claim concrete.
GPU kernel leaderboard
My batched dense Cholesky submission placed 6th of 107 entries on GPU MODE’s B200 leaderboard, with a 308.315 µs geometric-mean runtime across 15 benchmark shapes. A later natural-gradient audit passed 3 of 8 additional shapes, so I treat this as a leaderboard placement rather than a universal Cholesky result.
Open the artifact ↗Independent validation
Models propose compact Python programs. A separate deterministic validator searches for counterexamples before accepting a solution.
Open the artifact ↗Reasoning-model RL
A verifier-free dense reward that produced a four-seed SmolLM win over permuted reward and a full Qwen policy-learning replication on GSM8K, with majority and domain-transfer non-wins retained in the evidence trail.
Open the artifact ↗High-recall retrieval
A hackathon system that proposed real-time knowledge updates from X posts, with the event validator deciding which candidates were admissible.
Open the artifact ↗Interactive research note
A Marimo notebook for exploring over-mixing, sink tokens, and distributed memory behavior in attention.
Open the artifact ↗Good problems
Sparse structure, model forensics, optimization, verifiers, and experiments where a failed gate teaches us something are all fair game. If that sounds adjacent to your problem, send a note.
ajinkya@thepursuits.xyz