Notes from
the work.
Short stories about the problem I found, the choices I made, and what changed because of the work.
How I benchmarked NVIDIA Sirius and fixed two GPU bugs
I set out to compare a GPU-native SQL engine with DuckDB. The benchmark exposed two bugs first, so I fixed them upstream before measuring the performance.
- ✓113 join-heavy queries over 74M rows
- ✓6× average cache-warm speedup
- ✓Two fixes merged into NVIDIA Sirius
How I separated untrusted code from a Gradescope autograder
I found student notebooks running beside hidden tests and grade-writing credentials, then redesigned the grader so evaluation and publication happened across a real operating-system boundary.
- ✓Protected hidden tests and grade credentials
- ✓Separated trusted and untrusted execution
- ✓Preserved the existing grading workflow
How I compared Linux EEVDF and CFS under CPU contention
I used controlled workloads and eBPF traces to understand when Linux's newer scheduler improves responsiveness, and when longer, uninterrupted CPU time matters more.
- ✓4–6× lower p99.9 wakeup latency with EEVDF
- ✓Up to 65% higher Redis GET throughput
- ✓eBPF traces connected behavior to results
How I brought Textual apps to SSH with Textish
I liked how Charmbracelet's Wish made terminal apps available over SSH, but Python's Textual ecosystem had no equivalent. I built Textish to fill that gap.
- ✓One-line API for serving Textual apps over SSH
- ✓Four concurrency models explored
- ✓Fresh app state for every connection
How I built an AI-powered Slack bot to answer student questions
Students were already asking assignment questions in Slack, so I built a Gemini-powered assistant that could answer where the conversation was happening, with the course material needed to stay relevant.
- ✓Answers grounded in assignment manuals
- ✓Thread-aware conversations with image context
- ✓Four Slack-native commands