Work
Selected projects and research from our practice.
A Consent Layer for Cross-User Agent Interaction
Active research into the missing governance layer for AI agent communication — how agents representing different users enforce boundaries, propagate consent, and negotiate competing interests. Concept architecture paper in progress.
Evaluating Behavioral Protocols in an AI Knowledge System
A structured eval that diagnosed why an AI character reliably read its context files but never updated them at session end — then tested a fix and shipped a patch in one session. Demonstrates a baseline/diff methodology for objectively evaluating AI behavioral compliance.
What Text Evals Miss: Evaluating Speech-Native Models
Research brief on safety evaluation for full-duplex, speech-to-speech models — where the risk lives in timing, tone, and overlap that a transcript throws away. The questions we think evaluation has to answer.
Personalized AI Collaboration Framework
Designed a methodology for creating AI collaborators tailored to individual cognitive styles — moving AI from generic assistant to genuine teammate. The flagship implementation, Pemtu, is an AI project manager optimized for ADHD-aligned workflows and startup-speed execution.
AI Context & Memory System
Designed a lightweight DSL that solves AI's statelessness problem — enabling natural conversation continuity without database overhead or wasted tokens. Achieved a 23% improvement in continuity scores during user testing.