ALANCAI

L.I.O.N. is a hybrid reef-monitoring platform that runs underwater footage through YOLO-based detection pipelines to flag invasive species (lionfish, crown-of-thorns) and reef-health indicators, with annotated overlays and exportable detection data. Built with a Next.js front end, a detector-routing API gateway, hosted Roboflow inference, and a remote FastAPI service for heavier models.

Scio.ly is a free Science Olympiad practice platform serving a question bank built from 4,000+ past tests, where questions were extracted from raw PDFs using Gemini 3, then normalized and indexed across 23 Division C events with both MCQ and FRQ support. Built on Next.js (App Router handling frontend and backend via route handlers + tRPC) with CockroachDB via Drizzle ORM for persistence, Vercel KV for caching, and Supabase/Google OAuth for auth. Gemini 2.0 powers FRQ grading, on-demand explanation generation, and report processing. Built with TypeScript, Next.js, CockroachDB, Gemini 2.0/3, and Ably.

Tremora is a wrist-worn wearable that continuously monitors Parkinson's tremors and scores severity in real time against a proxy of the clinical MDS-UPDRS scale, replacing the ~1 hour of episodic clinical data patients get per year with 24/7 longitudinal monitoring. An ESP32 reads a MPU-6050 IMU at 200Hz, applying a per-axis IIR high-pass filter and a 256-sample Hamming-windowed FFT to extract tremor frequency and amplitude in the 3–12Hz band. A Random Forest classifier trained on PhysioNet recordings runs in parallel, and a physical medication-timing button timestamps doses so a Chart.js dashboard can plot severity before and after each one. Built with C++/Arduino firmware, ESP32, FFT/IIR signal processing, scikit-learn, and a vanilla-JS dashboard.

Provide is a free, public web map that helps North Carolina residents find SNAP retailers, food pantries, free meal sites, and WIC clinics near any address or ZIP code, overlaying USDA food-access data and NC transit routes to show both resources and structural barriers at once. Built as a no-backend static site, it queries live public APIs (USDA FNS, NC 211, Feeding America) at search time with bundled JSON caches as CORS fallbacks and Nominatim geocoding. Node maintenance scripts refresh each dataset, parsing the NC DHHS WIC directory straight from PDF, deriving statewide GTFS transit, and shading 2,195 census tracts to flag 353 food deserts. Built with vanilla JS, Leaflet, D3/TopoJSON, public civic APIs, and Node data-pipeline scripts.

Haku is a debate card search engine that indexes ~3.5 million pieces of competitive debate evidence from openCaselist (2013–2026), letting debaters filter by year, event, school, team, side, tournament, and round. A Rust ingest engine parses raw DOCX evidence into structured cards, a Python layer handles dedupe and indexing, and a staged pipeline imports everything into Meilisearch with rich card payloads checkpointed to SQLite for fast hydration at serve time. The system also supports paid collaborative team workspaces, shared libraries, and case-building tools. Built with Rust, Python, TypeScript, Next.js, Meilisearch, and SQLite/Postgres.