Projects

Architected and trained a custom 162M parameter conversational transformer model from scratch using JAX and Equinox. Optimized inference by compiling the entire autoregressive sampling loop into a single JAX JIT graph, achieving sub-second generation on serverless GPUs (via Modal). Integrated a FastAPI gateway with DuckDuckGo Web Search RAG to connect the model to real-time search results.

JAX, Equinox, Modal, FastAPI, RAG, Python

Built a cryptographically secure polling platform with Solana blockchain integration, supporting multi-wallet authentication (Phantom, Solflare, Backpack) and AI-powered fraud detection. Developed using React, TypeScript, Rust (Anchor framework), and Python FastAPI.

Solana, React, Rust, AI, Web3

Architected high-performance multi-tenant serving engine enabling 100+ concurrent AI coding agents with near-zero latency. Built Rust daemon with memory-mapped file system achieving 5-30x speedup over traditional shell-based search.

Rust, Python, IPC, vLLM

Constructed portable Pokédex device using Raspberry Pi 4 achieving real-time Gen 1 Pokémon recognition with 92.4% accuracy. Trained lightweight MobileNet CNN for edge inference and integrated LLM-powered backend.

Python, OpenCV, PyTorch, React.js

Developed full-stack music platform with Spotify API integration. Implemented emotion-based ML recommendation engine, increasing average user listening time by 18% through personalized content delivery.

React.js, Node.js, Spotify API, ML

Developed ML model using Python and LSTM algorithms that captures temporal stock patterns with 85% accuracy for predictive analysis.

Python, TensorFlow, LSTM, NLP

Created real-time F1 race winner prediction system using scikit-learn models trained on historical performance data. Built Flask backend with responsive Next.js frontend.

scikit-learn, Flask, Next.js