Write-ups on my research experiments and engineering projects.
A pointer-identity bug hunt in vLLM's RL weight-reload path (PR #48438, under review). The obvious A/B test passes on the broken code. An instrumented probe shows 88/88 runtime tensors rebinding under captured CUDA graphs, then supplies the allocation pressure that hangs the GPU at 100% for ten minutes on a generation that took 1.2 s before the reload. With an audit of eight sibling sites (all since fixed and live-validated, two of them via a synthesized 2 MB checkpoint), two retractions of my own findings, and a registry-wide version of the test that caught the same bug class in a second kernel on its first run.
A from-scratch implementation of Modal's snapshot-restore model on top of gVisor, with the full substrate underneath: content-addressed image format, unified blob store, custom lazy FUSE, thin scheduler, and a Modal-shaped Python SDK. Real workload (resnet50 on a 661 KB JPEG): cold p50 774 ms, warm 125 ms; runsc restore itself was ~280 ms at every stage. The interesting parts are the things that didn't work the first time.
A multi-agent pipeline for generating production-shaped and cross-language code where frontier-model priors are weak. Matched-budget GRPO on synthetic tasks shows monotonic +8.9pp learning where the same setup on public OSS oscillates and ends at −1.3pp.
A paired RLVR experiment on Qwen3-4B with algorithm and training budget held constant; only the task domain varied. Designed to test whether RLVR cross-domain transfer is task-agnostic or task-dependent. The two checkpoints produce opposite-sign GPQA Diamond effects, evidence that task choice is a real design lever in a post-training mix.
Boolean circuit minimization (Espresso as verifier) improves GPQA Diamond by +3.5pp, in the same ballpark as NP-Engine's classical-NP results. Monotonic across checkpoints (+1.5 at step 50, +3.5 at step 250), arguing against pure sampling noise.
RNA inverse folding (ViennaRNA as verifier) on a domain with zero pretraining overlap: strong in-domain learning (30% → 52% perfect solve), but cross-domain transfer is −4.1pp on GPQA Diamond, opposite sign at matched algorithm, model, and budget.