I build retrieval and memory systems that let you ask “Where did that answer come from?” At MiQ I work on market analytics. In open source I work on a related question about software behavior. When a form loses a row or an agent repeats an action, which step broke the promise to the user?
IntelliRAG · Ask a question, open the cited passage, then follow the retrieval trace to see which evidence survived. The public demo starts with cited extracts. Open the lab · Read the evaluation
memoRABLE · Turn a document into six reusable memory blocks. Follow a block back to its source, then use that same material in an email or on a web page. Try the workspace · Explore the code
ThermoSense · How much does a model improve on “tomorrow equals today”? I tested that baseline over a 365-day window using the Open-Meteo archive. A regularised linear model reduced RMSE by about 8% on that window. The audit explains how the reference data limits that result. Open the dashboard · Read the forecast audit
- A retry should produce one artifact. In OpenFuse, a file could exist even when its checkpoint was lost. I made the retry recover the original file. OpenFuse #45
- Deleting one row should leave its neighbor intact. In React Hook Form, removing a parent row could unregister the nested fields of the row moving into its place. I fixed the interaction between removal and reindexing. React Hook Form #13828
- One tool definition could break every request. Hermes sent a schema constraint that llama.cpp could not compile. I removed the incompatible constraint while preserving the runtime length check. Hermes #131286
- A list should follow the page around it. Removing content above a virtual list could leave it displaying stale rows. I fixed the missed update in react-window. react-window #935


