Messed around
and found out.
Research spanning coding agents, systems, databases, healthcare, and applied machine learning.
Agentic Coding Reliability · NSF Cybersecurity-Funded
An NSF-funded study of whether software written by LLM coding agents is actually safe to ship. I rebuilt a classic UNIX reliability fuzzer into faster research tooling, tested a 12-program corpus of utilities including grep, make, and ptx, and root-caused real memory-safety and denial-of-service bugs.
- 12 programs · grep, make, ptx + more
- 4× faster evaluation tooling
- 10+ source defects investigated
- uutils ptx · 29× faster · 19% less peak memory
Awaaz-e-Sehat
A React Native mobile app deployed in hospitals for seven months, using Whisper fine-tuned with LoRA on local Urdu speech to create structured records and surface dangerous maternal-health conditions.
- React Native · 500+ records
- 96% accuracy
- 60% less documentation time
- 7%→40% detection · 300+ alerts
Designing Around Stigma
A WhatsApp LLM chatbot for menstrual-health education, grounded in expert-curated content to minimize hallucinations and evaluated in a low-resource setting.
- 403 deployment messages
- 2-week deployment
- AWS EC2 + DynamoDB
- ACM CHI 2026
Evaluating Linux Schedulers
An empirical comparison of Linux EEVDF and CFS across latency-sensitive and CPU-bound workloads, combining application-level measurements with eBPF traces under controlled contention.
- EEVDF · 4–6× lower p99.9 wakeup latency
- EEVDF · up to 65% higher Redis GET throughput
- CFS · 13–59% lower HTTP latency
- CFS · 10–15% higher 7-Zip throughput
VIDER Video Reconstruction
A model-agnostic streaming framework that sends sparse keyframes over long-haul links and reconstructs intermediate video frames near the viewer.
- 97.69% avg. bandwidth savings
- 0.83 avg. ViSiL similarity
- 28-video evaluation
- Beat H.264 CRF 50 on savings + similarity