Computer Engineering @ University of Washington · Class of 2028
I build machine learning, full-stack, and assistive technology that people actually need.
I'm a Computer Engineering student at the University of Washington who builds at the edge of hardware and intelligence — from low-level systems to ML models that draft sentences for non-verbal users, watch over Alzheimer's patients offline, and surface healthcare disparities hidden in the data.
| 🚀 Builds shipped | 🏆 Hackathon awards | 🎓 Graduating | 📍 Based in |
|---|---|---|---|
| 8 | 4 | 2028 | Seattle, WA |
| 🥇 Honor | 🔎 What it was |
|---|---|
| 🥇 1st Place — ConnectAble (2026) | Predictive AAC system for non-verbal users · 3-layer ML cascade |
| 🏅 4th Place — Linx, BeaverHacks 2026 (NVIDIA Track) | Offline local-first AI for Alzheimer's care on a ~$150 Raspberry Pi |
| 🎖️ Honorable Mention — DubsTech Datathon 2026 | Healthcare access inequity analysis · 93.7% ML accuracy |
| 🎖️ Honorable Mention — Sea Score (2025) | Sustainability app with zero-shot ML photo verification |
Languages
Frontend
Backend & Data
Machine Learning
Cloud, DevOps & Hardware
A personalized Augmentative and Alternative Communication system for non-verbal users. A 66-symbol board lets users compose sentences and speak them aloud, while the system learns their voice over time. Predictions run a 3-layer cascade — bigram lookup → ChromaDB vector search → Ollama
phi3fallback — and a nightly 2 AM job retrains continuously. Everything runs locally; phrases are stored in AES-256 encrypted SQLite. An agent tab classifies natural-language intent and dispatches tools.
-000000?style=flat-square&logo=ollama&logoColor=white){kind=icon}
Local-first care companion that runs entirely offline on a ~$150 Raspberry Pi stack (camera, ultrasonic sensor, 4-mic array). Sensor signals fuse into a rolling memory buffer reasoned over by NVIDIA Nemotron 3 via Ollama, while a parallel deterministic rule engine catches high-risk patterns (stove left on, extended inactivity) before inference — rules can raise risk but never lower it. Backend is six MCP microservices. Built in 24 hours by a team of 4.
Three complete ML pipelines surfacing healthcare inequities across 75 demographic subgroups. Benchmarked 13 algorithms to hit 93.7% accuracy on cost-barrier prediction, forecast 2025 care-delay trends (9.15% avg, 95% CI 7.93–10.36%), and cluster 11 high-confidence anomalies at a 0.815 silhouette score — revealing a post-COVID rebound in barriers and mental health as the only worsening category. ~90s total runtime.
Eco-conscious mobile app that rewards real-world sustainability actions. Users complete challenges, upload photo proof, earn points, and unlock rewards. Photo validation runs via zero-shot object detection (
Xenova/owlv2-base-patch16) — no retraining per challenge — alongside a passport, rewards marketplace, team leaderboard, and community feed.
-61DAFB?style=flat-square&logo=react){kind=icon}
On RSVP, a local LLM summarizes your profile into a 768-d embedding and embeds your intent. Matching blends profile + intent similarity (60/40) via pgvector cosine search over an HNSW index, applies a skill-diversity constraint, and generates a personalized icebreaker per match — all on-device, no cloud. iOS app surfaces a discovery feed, top-4 matches, bookmarks, and live status.
-0071E3?style=flat-square&logo=swift&logoColor=white){kind=icon}
Chrome extension + event-driven AWS backend that detects when someone is struggling online (repeated reloads, error pages, stuck forms, scam sites) and emails a caregiver a plain-English summary with a screenshot. At local score ≥10 it captures a tab and POSTs to API Gateway → Lambda writes to DynamoDB (30-day TTL) and S3. EventBridge re-scans every 15 min, and at aggregate score ≥5 publishes an SNS email with a presigned link.
Nudge — A mobile banking feature for Sound Credit Union that surfaces personalized deals matched to spending habits, tracks savings over time, and delivers GPT-4o-mini financial advice and projections, all inside the existing app.
SolarSave — A full-stack solar savings calculator. Enter any address; it geocodes via OpenCage and pulls real irradiance from the NREL Solar API to return estimated annual savings and CO₂ offset.
🔗 Explore every build with full write-ups at nipunsaini.com
╔══════════════════════════════════════════════════════════════╗
║ 📚 Math Tutor · King County Library System ║
║ → 50+ students mentored · measurable curriculum gains ║
╠══════════════════════════════════════════════════════════════╣
║ 🏥 Hospital Associate · Evergreen Hospital ║
║ → HIPAA-compliant records · clinical emergency response║
╚══════════════════════════════════════════════════════════════╝
