Projects

A vector memory engine written in Rust that works with Redis clients through the RESP protocol. Uses HNSW for similarity search, SIMD distance calculations, and a write-ahead log for recovery.

A PyTorch library that treats image features as a graph and learns to segment them using spectral methods. Includes custom graph operators and a differentiable eigensolver, with evaluation as a U-Net segmentation head on Cityscapes.

A CPU-based autograd engine in Rust for studying computation and memory use in neural networks. Combines a dynamic computation graph, matrix multiplication kernels, task scheduling, and a reusable pool for gradient memory.

An LLM research project for scientific reasoning, mathematics, and formal logic. Combines energy-based objectives with autoregressive generation to study iterative self-correction.

A real-time chat application with saved messages, online presence, typing indicators, and chat rooms. Uses Socket.IO for communication and MongoDB for message storage.

An Android app that groups visually similar photos. Extracts image features and clusters them on the device, keeping photos off external servers.

A simulator that generates training datasets from grid-based environments. Supports configurable physics, agent policies, and collision rules, with repeatable runs and batch export.

A system for detecting and collecting microplastics in water. Combines optical particle measurements, signal analysis, and acoustic filtering to identify and separate particles.