I build AI tools in Python and TypeScript, with a focus on agent memory, workflow controls, and evaluation. My projects include SQLite retrieval, MCP servers, task queues, model routing, and research pipelines that record failed ideas.
I am researching decoder-guided agent recovery using entropy measurements and recent agent activity. The idea is to help a supervisor choose a recovery action after a failure or an uncertainty spike. I have prototype measurement and intervention code, plus exploratory analyses. The decoder design still needs implementation and a controlled evaluation.
The notes include an experiment plan, a proposed hidden-state probe, and work completed so far. I have no validated result showing that the proposed decoder improves agent recovery.
The setup guide links the memory MCP installer, shared project context kit, and three skills for reviewing project work. It includes the tested versions and instructions for choosing the parts you need.
| Project | What you can review |
|---|---|
| Multi-Agent Memory MCP | SQLite FTS5 retrieval, a typed graph, bounded two-hop search, and a synthetic benchmark |
| Jarvis Control | Task queues, model routing, scoped memory, and typed authorization for AI workflows |
| Team memory kit | Portable project instructions, blank memory templates, and a context builder |
| Agentic Quant Operator | Research controls, paper ledgers, and recorded promotion refusals |
| Event Contracts | Python and Rust contracts, replay, risk checks, and paper execution |
| PMQS | Execution-aware backtests, settlement accounting, and evidence checks |
| CIFAR-10 project | A ten-layer CNN, training code, and eight-seed validation records |
| Casino Line Modeling | Sports-model calibration against market baselines, including negative results |
| Multi-agent system report | Observations of throughput, review capacity, and memory pressure in one system |
The memory benchmark uses synthetic data. The CNN result measures selected validation accuracy. The trading research does not establish an edge, and Jarvis Control is a development project without reported revenue. Each repository explains its test setup and evidence limits.
I use AI coding tools throughout my work. I focus on defining the task, checking the generated changes, and making results inspectable. The project notes describe additional work whose source stays private.
Contact: GitHub