I architect production AI systems—from autonomous desktop agents and financial analysis platforms to healthcare applications and developer tools. Not prototypes. Production.
I build systems that ship to production — not prototypes gathering dust on GitHub.
From autonomous AI agents with sandboxed execution to SEC filing analysis with vector search, prediction market trading systems, and healthcare platforms serving real patients — each project solves a real problem with battle-tested architecture.
Universal AI Agent
Sophisticated native desktop AI agent in Rust/Tauri 2 with hexagonal architecture across 17 crates and ~800 tests. Vision-capable LLM routing, extended thinking, MCP runtime, multi-provider failover (Anthropic, OpenAI, DeepSeek, Ollama, llama.cpp), and true cross-platform sandboxing (bwrap, Seatbelt, Job Objects). Local-first semantic memory with vector search + knowledge graph.
NPU Quantized Inference
Core operator contributions to AMD's official NPU programming framework. Built the complete W8A8 quantized inference pipeline — quantize (bf16 → i8), dequantize (i32 → bf16), and scale_i32_bf16 operators — enabling end-to-end INT8 GEMM on AMD NPUs without CPU round-trips. Validated against Llama 3.2 1B.
RAG Security Hardening
Security fix for AMD's official Ryzen AI agent framework (1.2k+ stars). Patched a critical pickle deserialization vulnerability in the RAG cache layer with SHA-256 integrity verification, size limits, automatic corruption detection, and graceful re-indexing. Backward-compatible migration path for existing caches.
Mental Health Platform
Therapy marketplace for Spain connecting patients with licensed psychologists. Real-time video sessions, AI-powered session summaries, Stripe payments. Fully GDPR/LOPD compliant with EU data residency.
Financial Intelligence Platform
SEC 10-K filing analysis and financial risk assessment. RAG pipelines with BGE-M3 embeddings, Qdrant vector search over 500+ companies, multi-dashboard analysis interface with quality scoring.
Prediction Market Trading
Algorithmic trading system for Polymarket and Kalshi. BTC 5-min sniper strategy, cross-platform arbitrage detection, whale/insider tracking, real-time WebSocket feeds, and Telegram alert integration.
Semantic File Search
Semantic file search powered by local NPU hardware. Hybrid cosine similarity and keyword matching, incremental indexing, PDF/Office document support, and an interactive fzf-like selection mode.
NPU Shell Copilot
Shell copilot powered by local NPU inference. Auto-fixes failed commands, explains complex pipelines, suggests solutions from natural language. Non-blocking background LLM calls, works with any OpenAI-compatible API.
MCP Server Pioneer
One of the earliest Model Context Protocol servers for bash execution. Built before the MCP ecosystem existed—pioneering the tooling that would become the standard for AI agent integration.
Real Estate CRM
Intelligent real estate CRM with RAG-powered compliance features. Full-stack application with Next.js frontend and PostgreSQL backend, deployed and serving live users.
Voice AI Agent
Real-time voice AI agent with full audio pipeline. Voice Activity Detection, Speech-to-Text, and Text-to-Speech running over WebSockets for seamless conversational AI experiences.
Interested in working together? Let's talk.