Training an LLM to use a Debugger
Post-training Qwen3-Coder to debug with breakpoints, stack inspection, and live evaluation
Optimizing NVFP4 Grouped GEMM on Blackwell – Worklog
Taking an NVFP4 Grouped GEMM kernel from 238 µs to 23.8 µs with TMA, warp specialization, and cluster multicast.
21 min
Flushable SSTables in Pebble
~60% perf improvement when ingested SSTs overlap with the memtable
6 min
Enforcing Robust Type Safety with ESLint
Surfacing ~250 type inconsistencies in Linear's type definitions
4 min
Shipping Linear Drafts
Making drafts a first-class entity throughout Linear
8 min
Concurrent Manual Compactions
Segmenting key space and parallelizing execution, 30% perf improvement
6 min