// ansh-verma.about() shipping @ Litt
i build agents that watch other
agents work, and the kernels
that make them fast.
ai/ml engineer · b.tech cs (ai/ml) · northcap '25 · 9.0 cgpa. currently architecting judge-as-a-service for multi-agent llm trace forensics. previously: gpt-2 from scratch in c++/cuda.
work[].json 4 entries · sorted by recency
01 eval-agent shipping ↗
judge-as-a-service · llm trace forensics
otelfastapiqdrantlangchainpydantic
p50 retrieval 420ms −92% ▁▁▁▂▁▃▂▄▃▅▄▆▄▅▆▆▇▆▇█
02 gpt-2 · cuda archived ↗
transformer kernels in raw c++/cuda
cudac++pytorch
tok/s @ 124M 42k +3.1× ▁▁▁▁▂▃▃▄▄▅▅▆▆▆▇▇▇███
03 dineleap live ↗
qr-ordering · live menu · realtime analytics
honoredismongozodcloudflare
order TTI 180ms −67% ▂▂▁▂▃▃▂▄▄▅▄▅▆▅▆▆▆▆█▆
04 browser-automation shipping ↗
vision + grounding on government portals
playwrightpytorchvlmfastapi
flow success 94% +41pp ▁▁▂▂▃▃▄▄▄▅▅▅▆▆▆▆▇▇▇█
project.timeline 12 months · 4 projects
eval-agent shipping
gpt-2 · cuda archived
dineleap live
browser-automation shipping
notes/ 6 entries
2026.05 how we built the eval system at litt eval →2026.04 why i write cuda kernels by hand systems →2026.03 judging an llm is a search problem eval →2026.02 on logging at the trace level, not the request level otel →2026.01 attention is a transpose systems →2025.12 a kernel journal · ch. 1 systems →