AI Systems Engineer · 7 Production Projects · 679+ Automated Tests
I design, build, operate, and manage production AI systems on local GPU infrastructure. I specify intent to machines with surgical precision, evaluate every output as if it has my name on it, decompose work into multi-agent systems, and diagnose failure patterns at root.
Aether — Local-First AI Companion · github.com/dbhavery/aether · Showcase
A companion that lives on your machine, remembers you across sessions, and never calls the cloud without your say-so. Seven-layer Rust architecture — L1 Interaction · L2 Memory · L3 Presence · L4 Router · L5 Policy · L6 Persona · L7 Trust UX. The policy gate (L5) is non-bypassable: every side effect goes through the approved execution path or the linter blocks the PR. Audit log is SHA-256 hash-chained and HMAC-SHA256 sealed — tamper-evident, not tamper-proof, and the doctrine word matters. 11 crates, 4 timing contracts (250 / 800 / 2000 / 4000 ms), 2 SQLite migrations, MIT. OSS preview, pre-0.1, active daily.
Rust Tauri Cargo Workspace pnpm SQLite HMAC-SHA256 Ollama ts-rs specta cargo-deny
Vault — Self-Hosted Cloud Library
52,400+ files, AI captioning, face clustering, semantic search across 85K+ embeddings, RAG chat. FastAPI backend, React frontend, React Native mobile with camera roll auto-upload. Google Photos and Drive sync.
Python FastAPI React TypeScript React Native ChromaDB BLIP InsightFace Peewee SQLite Tailscale
Forge — Desktop Image Generator & Editor
Dual-model architecture (FLUX.1-dev for quality, RealVisXL Lightning for speed). txt2img, img2img, inpainting. 144 tests.
Python PySide6 diffusers FLUX.1-dev RealVisXL CUDA
Herald — Autonomous Intelligence Pipeline
13 RSS feeds + web search, 6-dimensional scoring engine, Claude draft generation with self-review, Playwright auto-posting to LinkedIn. Analytics feedback loop. Zero human intervention. Runs daily.
Python Peewee Playwright Claude API Ollama Firebase
VoxType — Push-to-Talk Dictation
GPU-accelerated STT, LLM smart cleanup, paste at cursor. Lockfile coordination with Aether.
Python faster-whisper Ollama CUDA
Portfolio — 3D Interactive Showcase
Next.js, React Three Fiber, custom GLSL shaders, Rapier physics, GSAP scroll choreography. 6 WebGL scenes.
Next.js React TypeScript Three.js R3F GLSL GSAP Lenis Vercel
22 custom agent definitions, 22 skills, 6 MCP server integrations. Multi-agent decomposition with parallel subagents, verification loops, and atomic commits. Token routing across model tiers. Context architecture for retrieval. GPU memory management across competing workloads.
AI/ML Claude Gemini Ollama PyTorch CUDA ChromaDB HNSW RAG MCP diffusers
Voice & Vision Porcupine faster-whisper ElevenLabs Chatterbox ECAPA-TDNN LivePortrait MuseTalk BLIP InsightFace
Backend Python FastAPI asyncio WebSocket Peewee SQLite
Frontend React TypeScript Next.js Tailwind PySide6 React Native
3D/Graphics Three.js React Three Fiber GLSL Rapier GSAP
Infrastructure Tailscale Firebase Playwright Vercel
Master's Degree — In Progress
Email: dbhavery@gmail.com | LinkedIn: linkedin.com/in/dbhavery | Portfolio: dbhavery.dev