Senior AI Engineer · Rust developer
I design and build autonomous agent workflows and heavy automation where correctness, safety, and predictability are non-negotiable.
My work is agent orchestration, long-running LLM pipelines, retrieval and memory, and the deterministic control layers that keep autonomous behavior auditable and bounded — mostly in Rust and Python, with explicit guarantees around latency and failure modes. I tend not to build usual things: a temporal vector store, a behavioral layer with its non-goals written into the type system, a Bitcoin L2, threat research on the side.
A temporal vector store — time treated as a first-class axis of similarity, not an afterthought.
HNSW graphs with native temporal decay and importance weighting, over memory-mapped storage and lock-free concurrent operations.
A behavioral layer that turns human state into machine-consumable interaction constraints for LLMs.
23 interpretable axes; a pure behavioral layer that produces prompt context, not actions. Explicit non-goals: no action execution, no persuasion optimization, no hidden optimizers.
Bitcoin UTXO management and off-chain state for payment channels.
Merkle-based state proofs give O(log n) verification; deterministic signature ordering prevents replay. Built for provable, auditable state transitions.
Multi-agent graph intelligence — natural-language querying over agent-interaction graphs.
Schema-agnostic backend that evolves JSON schemas at runtime; NL translated into safety-constrained AQL (whitelisting, result caps). Refactored the LLM subsystem down ~84% while improving reliability.
A verification gate that commits only the claims a source actually supports.
Cross-family LLM cross-checking as a correctness layer: unsupported claims are dropped, not softened.
Reverse-engineered a credential-theft NPM package disguised as a recruiter take-home.
Traced obfuscated payloads through fake passport middleware into a staged execution pipeline; reported the package to NPM for takedown.