Projects

The layer that sits between a model and an agent that actually does something. Handles the parts that every agent needs but nobody wants to rewrite: persistent state, checkpointing so long runs can be interrupted and picked back up, provider-agnostic model access, and clean module boundaries so capabilities don't leak into each other. Works with OpenAI, Anthropic, and local models.

A read-only agent for understanding large codebases. The interesting part is the citation layer — every claim the model makes about the code is checked against the actual source, so hallucinated function names and invented behavior get caught before they reach the user. Uses a mix of exact search, structural search, and semantic search depending on what the agent is looking for.

A personal assistant capability that handles tasks, journaling, and reminders. Mainly a testbed for what useful long-term memory looks like in practice — searching past entries by meaning rather than keywords, keeping an audit trail of what the agent changed and why, and handling concurrent edits safely.

A voice interface for Jeeves that runs on-device — wake word, speech-to-text, and text-to-speech all happen locally, with no round-trip to the cloud. Uses the onboard NPU to keep the models fast enough to feel responsive. Part of the broader interest in pushing more of the agent stack onto local hardware.

Training neural networks to generate quantum states is hard because most outputs aren't physically valid. The trick here was baking the validity constraints directly into the architecture, so every sample the network produces is a legitimate quantum state — no rejection sampling, no post-hoc fixing.

Training models to tell whether a quantum state has useful entanglement, using far fewer measurements than the standard approaches need. Tested across a range of canonical states.

Symbolic math work on how entangled different multi-qubit network configurations are, and how well they'd hold up as a resource for quantum teleportation.