Selected work
Hover a photo to peek at another screen. Tap it to open the real interface.
Dual Hall-sensor homing on ESP32 to restore wheel zero after power loss. CAN (TWAI) motor control at 100 Hz for velocity and incremental positioning, hysteresis against EMI, and a diagnostic web UI that cut field setup time.
React + Python arm teleoperation over CAN: ACK/recovery, slider and gamepad controls with joint limits, disconnect handling, commanded-vs-feedback TX, HEALTH state machine, and a UART drive backup path. Merged into urc-teleoperation-2026.
Arduino UNO monitor for indoor gas: LCD1602 status, green/yellow/red LEDs, and a fan that turns on when the reading is harmful or dangerous. Built with Wenyan Chen, Ricardo Barron, Luke Nguyen, and Kevin Ngo. The DHT11 overheated during testing; the gas path, lights, and fan still ran.
Source-linked portrait prototype from public web material: username discovery, a SQLite corpus with crawls and citations, and LLM reports that stay grounded in saved evidence. Shipped backend contract tests and Playwright coverage, plus a production-build fix for empty Next.js routes.
iPhone-first PWA: photograph a product or type the details, then a FastAPI backend estimates opened/unopened dates, risk, and a recommended action. Camera capture, schema-validated analysis, and on-device saved results — educational estimates, not medical advice.
Laptop prototype that records a lecture locally, transcribes it with on-device Whisper, and turns the transcript into an editable to-do list. Audio, text, and tasks stay on the computer. Still in progress.
High-concurrency event platform for SJSU SCE: Kafka registration/waitlist pipelines, Redis locks under 10k+ write spikes, Mongo persistence, and a React calendar on sce.sjsu.edu with JWT auth, role visibility, scheduled publish, and closed-event blocking. 20+ PRs across SCEvents and Clark.
Telecom churn models on usage, tenure, and service features. Compared logistic regression, random forest, and gradient boosting; shipped the 79%-recall logistic pipeline through FastAPI with a saved scikit-learn preprocess/encode/scale/infer pipeline. Scaled experiment on 250k rows reached F1 0.72 and ROC-AUC 0.90.
Multi-agent research pipeline: crawl Wikipedia with Playwright, structure results, store in Redis, then serve predicted follow-ups with semantic KNN (RedisVL) plus LLM-filtered prefetch so warm queries skip a recrawl.
Paste a messy transcript, get summary, decisions, action items with owners/deadlines, blockers, and a copy-ready follow-up. FastAPI + NVIDIA Nemotron with schema validation; React/Vite UI.