Experimental AI architecture | local-first systems | bounded agent readiness
I build and document experimental AI systems focused on one practical question:
Before an assistant receives tools, memory writes or runtime authority, can it become more coherent, traceable, socially repairable and useful?
My current public work is centered on Lucie / Himadjin Lab, a local research showcase for cognitive routing, architectural memory, defensive-loop probes and human-validated assistance.
| Area | What I am exploring |
|---|---|
| Local AI architecture | Compact layers that improve routing, continuity and traceability around language models. |
| Pre-runtime readiness | Evaluating whether an assistant is stable enough before tool access or stronger authority. |
| Human-in-the-loop design | Keeping the operator in control while preserving useful initiative. |
| Conversational repair | Detecting loops, refuge questions, protocol leakage and poor social recovery. |
| Bounded exploration | Separating creative hypotheses, toy estimates and validated claims. |
Lucie is the visible conversational surface. Himadjin is described publicly as a compact local substrate for orientation, contextual routing, bounded scoring and experiment framing. Jarvis is the orchestration layer that keeps routes, traces and safety boundaries readable.
- Repository: lucie-himadjin-lab-public
- Demo: GitHub Pages showcase
- Release: v0.1.1 Visibility Pack
- Discussion: feedback on pre-runtime agent readiness
- Local first: private runtime and sensitive traces stay local.
- Operator governed: stronger permissions require explicit validation.
- Traceable before powerful: useful systems should explain what they know, what they infer and what remains unproven.
- Creative but bounded: invention and simulation are welcome, but they must remain separated from scientific validation.
- Socially mature assistance: a useful assistant should recover from correction without becoming evasive, defensive or cryptic.
This profile and the public repository do not claim artificial consciousness, subjective experience, scientific discovery, production security assurance or a released autonomous agent.
The public work is a research showcase: positioning, architecture, cleaned examples, benchmark scaffolds and explicit limitations.
- feedback on benchmark design;
- discussion with local AI and agent-evaluation builders;
- reviews of public wording and claim boundaries;
- collaboration around traceable, operator-governed assistant systems.
If the project resonates, the best entry point is the public discussion thread: