I work on LLM inference, evaluation and post-training, mostly measuring how serving systems actually behave and building tools around what I find.
Portfolio · LinkedIn · X · ORCID
- Reload or Recompute? (preprint): a break-even rule for when a resumed conversation should reload its KV cache instead of recomputing it, tested on L4, A10 and A10G with predictions frozen in advance.
- Leaderboard Noise: "deterministic" greedy decoding changed 2.0–4.6% of answers across GPUs and reversed a near-tied ranking.
- Grammar vs Bits (preprint): strict grammar decoding gave no clear recovery of what 4-bit quantization loses in tool calling; the output format mattered far more.
- Pixel Miser (write-up): post-trained Inkling 2 Small on Tinker to stop over-thinking. Output tokens fell from 3,109 to 74 and finish rate rose from 49% to 99%.
- Nightly (in progress): a coding agent that retrains on its own test-verified attempts each night, with hidden-test graders red-teamed against cheating.
- Kestrel: multi-tenant GPU compute platform on Kubernetes with quotas, GPU-second metering and autoscaling.
- kea: real-time incident root-cause engine. 1st place, OpenAI Codex Community Hackathon.
- AntiVibe: Claude Code skill that explains AI-written code. 1,100+ stars, Mintlify OSS Program.
- SentinelLM: LLM proxy with a concurrent safety-evaluation pipeline.
- Google Summer of Code 2026, RUXAILAB: built the Focus Group Management and Analysis Module (24 merged PRs).
- Koncepts: Product Developer Intern, 165+ PRs to an AI EdTech platform serving 200K+ learners.
- India Builds with Claude: 1 of 6 presenters selected from 2,300+ applicants.