Agents, infrastructure, and full-stack applications — built and running in production.
Telegram bot bridging directly to Claude AI. Send a message from anywhere — Claude responds with full tool access including web browsing. Monitored via live dashboard.
AI agent governance and audit layer. Every action the AI agent takes is logged, risk-scored, and auditable. Human-in-the-loop approval for sensitive operations and full decision history.
Multi-agent LLM system for desktop automation. Local-first with CrewAI orchestration and Ollama inference, with no cloud required.
Semantic knowledge store and memory bridge for Claude Code. Extracts facts from session logs into a RAG database and surfaces relevant context at session start.
Dell i5-4690K tower repurposed as a dedicated dev/data lab. Runs JupyterLab, Docker, and PostgreSQL. Connected via Tailscale. Home for web scrapers, data pipelines, and ML experiments.
One-shot bootstrapper for a private local AI stack: no cloud, no subscriptions. Installs Ollama, pulls Qwen2.5-1.5B, and sets up Open WebUI as a systemd service. Runs on any Ubuntu/Debian machine, CPU-only.
Local AI backend designed for a Raspberry Pi 4 (ARM64). FastAPI + llama-cpp-python inference server with streaming responses. Runs offline, no cloud required. Goal: a persistent home AI that remembers context and never phones home.
Local LLM embedded in physical objects: a bear, a photo frame, a card. No internet, no subscription. Fine-tuned on personality data using QLoRA + voice cloning with Coqui TTS. Privacy-first memory preservation for loved ones.
Self-contained LoRA/QLoRA fine-tuning toolkit for open LLMs. Config-driven training, BLEU/ROUGE/perplexity evaluation, evolutionary hyperparameter search, and a FastAPI proposal review dashboard.