Most AI forgets. I build the kind that remembers what worked — and ships the code that proves it.
Founding / Staff AI engineer. I build self-improving AI systems on the hard primitives — late-interaction retrieval, agent memory, and reinforcement-style credit assignment — and ship the code that proves them.
Most AI forgets: a model scores 90% today and 90% tomorrow; what worked, what failed, what to avoid is generated and thrown away. I work on the opposite — systems whose judgment compounds from reality, getting sharper with use because outcomes are scored and bound to the next decision. That thesis is Accreted Intelligence .
| Repo | What it proves |
|---|---|
| accreted-intelligence | The thesis + whitepaper: a Recursive Language Model over late-interaction scored-token memory, where the model is a replaceable processor and judgment lives in scored state. |
| maxsim-rs | ColBERT-style late-interaction MaxSim, in clean zero-dependency Rust — the retrieval primitive, done right. |
| colpali-retrieve | Multimodal late interaction — a text query retrieves over document-page images by patch-level MaxSim (ColPali-style). |
| scored-rerank | Ranking that learns from outcomes — Beta-Bernoulli posteriors + Thompson sampling rank by what reality confirmed, not just base similarity. |
| mcp-retrieve | An MCP server exposing late-interaction retrieval as an agent tool — retrieval where agents actually consume it. |
Core: late interaction (ColBERT · ColPali · MaxSim) · Recursive Language Models · RL credit assignment · agent memory · retrieval / RAG systems
Tools: Rust · Python · MCP · DSPy · PyTorch · SQLite · browser automation
Exploring Staff / Principal / Founding AI Engineer roles — remote from GMT−3, full US/EU-hours overlap, contractor-ready — in retrieval, agent memory, and RL-for-LLMs. CV on request.
"Reasoning is getting cheaper. Judgment is not. I build for the part that compounds."