I'm a scientist building software and a generalist by nature. Most of these projects started with something I wanted to make less frustrating: picking up a conversation, making sense of model behavior, starting a script, or getting a plate of data into a form I could actually use.

All of it was built with AI. Claude and ChatGPT are where I work a problem out; Claude Code and Codex write the implementation. Every project carries the same build credit because it is the honest one: both agents have been through every repository. Where the division of labour is worth knowing, the project page says so in its own words. Eighteen months ago I had never written code and had never used an AI model. These are grouped by subsystem rather than by date, but they were all built inside that window.

The work grew from there. Some projects are tools I use, some are experiments, and some are a closer look at a useful piece of a larger system.

A few places to start

Context and memory

How conversations become something useful for the next conversation.

• Context and Memory Workspace — my private workspace for chat, a generated wiki, search, and project memory.
• Attractor — Adaptive Context and Personalization — a changing map of how I work, and a way to turn it into context. The next part is choosing what to deliver for a particular interaction.
• Attractor — Knowledge Graph — a concept graph that lets me see connections across months of conversations.

Evals and interpretability

Tools for asking better questions about what the models are doing.

• Multi-Provider LLM Orchestrator — models from different providers working through a shared reading-room workflow, with an evaluation harness for comparing prompts.
• Attractor — Context Research — using the same Attractor mechanism to compare models and investigate recurring conversation patterns.

Attractor — Adaptive Context and Personalization describes Attractor's memory and context-delivery tool; Attractor — Context Research uses the same tool as a research instrument.

Developer tools and workflows

• Bootwitch CLI Scaffold — a local project starter with guided setup, script generation, and documentation that follows the code. The main project page includes both recorded demonstrations.

Interface

• 96-Well Plate Analyzer — a browser tool built around the plate-analysis work I used to do at the bench.
• Bootwitch Portfolio: UI Evolution — how this terminal-style portfolio found its visual identity.
• Guided Portfolio Conversation — how the assistant gained bounded continuity, a real fresh start, and safe visitor exports.

Routing and orchestration

• Model Routing and Job Orchestration — a closer look at how Context and Memory Workspace and Multi-Provider LLM Orchestrator choose models and move work through their systems.
• Bootwitch Doctor — How I Work with Agents — a multi-agent architecture repair experiment built around backward tracing, clear ownership, and independent verification.

Observability and subsystems

• Systems Logging and Telemetry — the parts that tell me what answered, what happened, and what needs attention.

Routing and observability are pieces of the larger projects, with their own pages so there is room to explain them.

A familiar way through each project

README is the introduction. ARCHITECTURE shows how it fits together. TECHNICAL gets into the interesting details. source contains selected scripts and excerpts, with context for what you're looking at.

Diagrams sit beside the explanations they support. Demos and research plans stay with their projects when they add something worth exploring. You can start with the overview and follow whatever catches your interest.