AI engineer focused on agentic systems — multi-agent orchestration, local-first RAG, and the infrastructure that makes LLM agents actually reliable. I build across the stack: Python backends, LangGraph pipelines, React/Next.js frontends, and the occasional Rust tool.
- 🔭 Now: neuro-symbolic RAG, multi-agent trading systems, cross-lingual speech ML
- 🌱 Practicing: Harness Engineering — treating the agent's environment as the product
- 💬 Ask me about: RAG, LangGraph, MCP servers, multi-agent design
"When things fail, don't swap the model — fix the harness."
A harness is everything outside the model weights. Same model, different harness, fundamentally different results — so I engineer five subsystems around every agent I run:
| Subsystem | What I ship |
|---|---|
| Instructions | Short, routing-oriented AGENTS.md — a map, not a manual |
| Environment | init.sh, lockfiles, ≤3-min rebuild from scratch |
| State | progress.md + git checkpoints — the repo is the spec |
| Execution | WIP=1: one feature active, next unlocks only after E2E passes |
| Feedback | Executable definition of done — verification decides, not agent confidence |
objective (AGENTS.md) → init → run task ⟲ runtime feedback
→ verify: lint → tests → full E2E (fail → loop back)
→ clean handoff (progress recorded, junk deleted, git green)
Mantras I work by: done = verification passed, not "code written" · someone in the crew must not believe you · do less but finish · give the agent a methodology, not a task.
|
Neuro-symbolic RAG agent, fully local. Hybrid memory (vector + knowledge graph), HyDE retrieval, self-correcting web fallback.
|
Four agents — research, write, review, publish — that ship finished tech articles to Dev.to autonomously.
|
|
Cross-lingual Parkinson's detection from voice. 88-feature eGeMAPS pipeline, 7-model evaluation across Spanish/English/Italian datasets.
|
Multi-agent trading platform where 8 investor personas debate every trade before execution.
|
🎵 SymphonyAI-powered terminal music player for developers who never leave the shell.
|
MCP server giving Claude full read/write access to an Obsidian vault — notes as agent memory.
|
+ LangGraph · DSPy · Ollama · MCP · Neo4j · pgvector
