A closer look
at the projects.
Each project follows the constraint, the tradeoff, and what I learned once the system met reality.
Userspace TCP/IP Stack
I built a dependency-free TCP/IP stack in Rust, then drove it through Linux TUN/TAP to prove Ethernet, ARP, IPv4, ICMP, UDP, and TCP interoperability end to end.
Impact7.6 Gbps read · 4.6 Gbps write through Linux TAP · 1,024 concurrent TCP connections · 28 protocol + Linux interoperability checks
TinyKV Distributed Store
I built a durable C++ key-value store that combines an LSM engine, key-value separation, crash recovery, and Raft consensus behind a small API.
Impact33,800 replicated writes/s · 8 clients · 12× faster writes for 16 KiB values · 65× throughput gain from group commit
SmolLM Alignment Lab
I implemented a 135M-parameter decoder-only language model in PyTorch, matched the reference model, then added optimized decoding, LoRA fine-tuning, and DPO preference alignment.
Impact100% top-1 reference agreement · 8.50× faster decode throughput · 0.73% trainable parameters with LoRA
Second Time Around
I built a multi-vendor marketplace for buying, auctioning, and donating pre-loved items, with seller dashboards and real-time commerce updates.
ImpactDirect sales + auctions + donations · Real-time bid and order notifications · Seller analytics and ratings
Gemini CLI History Search
I rebuilt Ctrl+R reverse search in Google's Gemini CLI as a reusable hook, so shell history and prompt history share one incremental search.
ImpactUsed by 1M+ developers · 2 PRs merged upstream · Debounced, cached incremental search
NVIDIA Sirius DB
I added the libcudf string aggregation Sirius was missing, then fixed the GPU correctness bugs the benchmark suite exposed.
Impact1K+ GitHub stars · 2 PRs merged into the query engine · Full industry benchmark suite unblocked
CS 220 AI Learning Infrastructure
CS 220 serves more than 900 students, and deadline days turn small inefficiencies into operational problems. I led a 12-person team and rebuilt the tutor, grader, and observability around the moments the course was under the most pressure.
Impact900+ students on autoscaling Azure infrastructure · 80% lower LLM inference cost · 12× faster grading · 12K+ submissions
Awaaz-e-Sehat
Paper charting was consuming time an understaffed maternity hospital did not have. I shipped the React Native app end to end, frontend and backend, working with doctors to design a voice-first workflow for noisy, code-switched Urdu and Punjabi speech, then carried it through a seven-month deployment.
Impact7-month live hospital deployment · Speech error rate · 46% → 13% · Structured records · 96% field accuracy
AI-Generated Code Reliability
I built a harness for AI coding agents to reimplement 10 Linux utilities, then built the compiler and fuzzing systems needed to measure whether the results were actually reliable. A Go-based tool traces C source into the linked binary, while the legacy fuzzer I rebuilt runs concurrently, resumes from checkpoints, and makes testing faster.
Impact10 utilities · 98% of upstream tests passed · Go-based tool · Go fuzzer · concurrent + resumable
Textish
I liked how Charmbracelet’s Wish could turn a Go terminal app into an SSH experience, but I could not find that bridge for Python’s Textual ecosystem. So I created Textish and remain its active maintainer: import a Textual app, call serve, and let anyone open it with an ordinary SSH client.
Impact3,000+ PyPI downloads · One shared Python interpreter · Real SSH + PTY load harness
Linux Scheduler Performance: EEVDF vs. CFS
I benchmarked EEVDF against CFS on 40-CPU bare metal and used eBPF traces to explain why neither scheduler wins every workload.
ImpactEEVDF · 4–6× lower p99.9 wakeup latency · EEVDF · up to 65% higher Redis GET throughput · CFS · 13–59% lower HTTP latency
VIDER Generative Video Reconstruction
I led a model-agnostic streaming framework that sends sparse keyframes and reconstructs the missing video with GPU compute near the viewer.
Impact97.69% avg. bandwidth savings · 0.83 avg. ViSiL similarity · 28-video evaluation
WhatsApp Menstrual Health AI
I built and field-evaluated a WhatsApp health assistant that grounds its answers in expert-curated menstrual-health content instead of model memory.
ImpactACM CHI 2026 publication · 2-week field deployment · Expert-grounded RAG
Food Crisis Forecasting with Urdu News
I led a Pakistan-focused food-crisis forecasting pipeline built on Urdu news and district weather data, reproducing and extending a Science Advances study.
Impact94% IPC prediction accuracy · 90× faster RF training · 162,653 Urdu articles
Slack AI Course Bot
I built an asynchronous Python Slack assistant for CS 382, LUMS's network-centric computing course, grounding Gemini in the course's own materials so students get help without leaving Slack.
Impact4 Slack commands · Assignment-grounded Gemini · Thread-aware responses
BookSwitch Browser Extension
I built and published a cross-browser TypeScript extension that puts the first nine sites on your bookmarks bar one keystroke away.
ImpactPublished on Chrome + Firefox stores · One codebase · two browsers · Alt+1–9 bookmark jumps
Meta Production Engineering Fellowship
Selected as 1 of 100 fellows for Meta's 12-week production-engineering curriculum with Major League Hacking, where I learned directly from Meta engineers and applied those lessons to real services. I kept a Flask/MySQL service stable on a memory-constrained VPS, shipped a second Go/SQLite service on its own host, and automated the release and monitoring path around both.
Impact1 of 100 fellows selected · 80% faster incident detection · Deploys cut from hours to minutes