01 / 03 · 2025–present
Orealis Labs
A deterministic control plane for long-horizon AI agents. It replaces unstructured LLM memory with an auditable, typed causal graph.
- Role
- Founder · AI Systems Research & Architecture
- Years
- 2025–present
- Prototype
- Orí, an AI companion
The problem
Language models optimize for a coherent conversation, not for the user's actual outcome. Without an external check, agent memory drifts into context loss, hallucinated relationships and sycophancy, and retrieval costs grow with the length of the transcript.
- Natural language in
- LLM codec proposes a lattice coordinate
- Lattice type-checks it for plausibility
- Unconfirmed, it waits as a hypothesis
- User confirms or corrects
- Placement locks
What I'm building
- 01An LLM-as-codec architecture. The model only translates natural language into exact coordinates on a fixed functional lattice. Deterministic graph traversal does the reasoning, prioritization and safety gating. No vector RAG.
- 02Provenance on every memory. Each assertion cites the immutable conversation turn that grounded it. When a user contests an inference, it is invalidated at that exact point.
- 03Expectation contracts. Every proposed resolution becomes a measurable prediction (condition, outcome, horizon) and is checked later. If reality misses, the diagnosis reopens.
- 04Precondition gating. The system cannot propose a route the user is structurally or financially blocked from taking.
- 05A breadth-to-precision retrieval ladder summary→detail→verbatim citation Token cost scales with relevance, not with transcript length.
Outcome
- A working prototype, Orí: an AI companion whose memory is a typed, causal model of what a person cares about, running today.
- Every prediction breaks down into named quantities you can inspect and correct.
- Design property The same coordinate system works as a privacy-preserving join key across users, for population-scale evidence.
“This system's mistake is a named edge you can fix.”
