Projects

SoulMate

Tinder-style sneaker discovery app with a personalized recommendation engine.

ReactTypeScriptFastAPIPythonSupabaseVite

Swipe-based sneaker feed with a FastAPI backend; auth uses Supabase's public JWKS endpoint so the server never touches a shared secret
Recommendation engine tracks a 7-dimensional taste vector per user, updated on every swipe via cosine similarity
Feed refreshes silently in the background after each swipe so there's no loading flash; added pointer event guards to fix a double-swipe bug

Social platform for sharing travel memories as photo + music pairings.

TypeScriptNext.jsPostgreSQLVercel BlobRedisSpotify API

Full-stack app using Next.js 15 App Router and Server Actions with custom auth and end-to-end type safety
Each post pairs a photo with a Spotify track and OpenStreetMap location to capture a place's vibe
Serverless infrastructure on Neon (Postgres), Vercel Blob for media storage, and Upstash Redis for rate limiting

Digital notebook for tracking daily habits and long-term goals.

Next.jsReactSupabaseTailwind CSSTypeScript

Journaling app with a custom book-style UI, dark mode, and Supabase-backed auth and storage
Goal planning with automated roadmap suggestions and visual habit streak analytics

Dual-axis solar panel tracker on an ESP32 that follows the sun in real time.

ESP32C++ArduinoOLED DisplayServo Motors

Four LDR sensors drive two servo motors to keep the panel aligned with the sun throughout the day
Adaptive control algorithm with noise buffering to smooth out motion under patchy light conditions
Live sensor readings and servo angles shown on a 0.96" I2C OLED display

Content-filtering pipeline for text-to-video prompts using an LLM ensemble.

PythonDeepSeekPhi-4LLM EnsembleAI Safety

Ensemble of Phi-4 and DeepSeek-R1 to flag unsafe text-to-video prompts before generation
Evaluated on 432 prompts; found a clear gap between non-temporal (93% accuracy) and temporal risk detection (61%), pointing to a model blind spot worth addressing

ViT model for hyperspectral image classification.

PythonPyTorchSciPyScikit-Learn

Adapted Vision Transformer for hyperspectral data using spectral-spatial patch embeddings; hit 95% accuracy on the Indian Pines benchmark