Explore my complete portfolio - from production AI systems to open-source tools
Showing 15 projects
MLOps framework unifying multiple fine-tuning ecosystems with automatic backend selection, AWS SageMaker integration, and significant cost reduction through spot instances.
Multimodal AI assistant that powered a 200K-user experiment, enabling natural language product discovery, offers, and rewards optimization on iOS.
Local-first portability for coding-agent sessions. Export a session from Claude Code, Codex, or Hermes and resume it in another harness, with every lossy conversion disclosed before anything is written.
A turn-governed communication layer for independent AI agents, with authenticated pairing, durable transcripts, and MCP and SDK access.
An open-source Rust CLI/TUI tool for developer-aware disk cleanup that scans for reclaimable space from build caches, AI models, package managers, and IDE artifacts.
A free, open-source iOS app for AI-powered task management with Focus Mode, Quick Capture, two-way Apple Calendar sync, and smart "Next Up" recommendations.
Multi-agent system for predicting brands and categories from ambiguous product descriptions. Mapped 15K+ products with 88% accuracy and became the foundational framework adopted for other catalog automation projects.
Production-ready AI context management using semantic similarity - 42% cost reduction with <12ms filtering. Features LangGraph workflows, FastAPI, PostgreSQL with pgvector.
LangGraph-based 3-agent workflow replacing double-annotation with LLM annotation at 80%+ accuracy, reducing total annotation workload by 40% for ML training data.
Autonomous robots designed and built for the ABU Robocon 2018 international competition. Features advanced control systems and embedded programming.
A 3D sound source localization system using an 8-microphone cubical mesh for locating disaster victims. Achieved 95% accuracy within 1.5m range using GCC-PHAT algorithm and CNN-based denoising.
Deciphering Earth's and Lunar materials with advanced machine learning. Uses deep learning and feature engineering for analyzing spectral signatures.
LLM-powered system to automatically flag incorrect product matches at 95% recall, protecting $800K in GMV annually and reducing daily audit effort from 8 hours to 15 minutes.
CLIP-inspired pipeline combining image and time-series encoders to predict sales patterns for new products before launch, using visual similarity to existing products with known sales history.
A security solution for anomalous activity detection in a network.
