projects
Tooling, instrumentation, and concept work for AI-native research.
A common thread runs through these: the runtime that consumes a tool can produce evidence about it. Hooks first. Configuration follows. Research stops being a separate motion and becomes a continuous read of a live system. The same instrumentation that surfaces evidence about user need surfaces evidence about how the system can mislead or be abused, intentionally or otherwise — designed for, not bolted on.
Smart Skills Platform
shipped · may 2026Researcher + builder
An evaluation and curation system for an internal library of agentic workflows. Failed prompts cluster into evidence; curators promote drafts on signal; skills are demoted when sources depreciate; new development is directed at the gaps the runtime surfaces. Two MCPs, project hooks across Claude Code and Cursor, threading + clustering backend.
User Research Slack Helper
in productionResearcher + builder
Perimeter-deployed Slack helper — Claude Code service account, frontier model, in-perimeter — with least-privilege tool access, human-in-the-loop fallback for out-of-scope queries, and full interaction telemetry. Anonymisation before ingest, permission-mirrored access; iterates through configuration of tools, MCPs, and skills, not raw model capability.
Inscryption Agent
in progressPersonal
Personal project building an RL agent for the deckbuilder Inscryption (Kaycee's Mod). Validated Rust simulator complete and surfacing actionable play patterns; warm-start training data run done; agent training next. The most transferable artefact so far: a framework for parallel agentic workflows.
Intersectional Decision-Support Tool
concept · scopingPersonal
Concept: decision-support tool grounded in intersectional analysis (Crenshaw, 1989; Hill Collins). Recognises that interlocking systems of power compound multiplicatively on individual experience to produce qualitatively distinct user contexts; designs at those intersections to produce disproportionate inclusive uplift while surfacing risks of harm that flat user models obscure.