I study how languages work and build tools from what I find.
Portfolio · Lab · LinkedIn · peleke@peleke.me
Can agents learn from their own performance?
Most systems marketed as "learning" are persisting context. Remembering, not adapting. I'm building the instrumentation to tell the difference: a five-layer architecture where learning can be observed and measured.
The full lab page has the system diagram, hypotheses under investigation, and current project status.
| Layer | Project | What It Does |
|---|---|---|
| 01 · Knowledge | qortex | Knowledge graph ingestion with typed edges and rule projection |
| 02 · Learning | buildlog / openclaw | Thompson Sampling over prompt components to measure what helps |
| 03 · Nervous System | cadence | Typed signal bus for ambient agency with temporal routing |
| 04 · Runtime | openclaw + sandbox | Agent runtime with process-level isolation via Lima VM |
| 05 · Interoception | interoception | Internal coherence monitoring via conservation quantity tracking |
| Project | What It Is |
|---|---|
| Interlinear | Language tutor that distinguishes a typo from a conceptual gap. 5 error types, different interventions. |
| Swae OS | Health platform with federated backend. GraphQL federation via Hive Gateway. |
| LinWheel | Paste a transcript, get 7 publish-ready posts. Including the articles LinkedIn's API blocks. |
| LangLine | Mobile-first terminal for developers. SSH + tmux on your phone. |
| ComfyUI MCP | 35+ MCP tools bridging Claude to distributed GPU compute. |
| Graphix | 201 MCP tools exploring the structural primitives of visual storytelling. |
Princeton CS · Principal Engineer @ edX/2U · AI Systems Fellow @ Overclock · Partner @ Endstation LLC