Inside LightBI’s 5.6 MB Micro Brain
No embedding API and no runtime LLM call. MB combines BM25, TF-IDF/LSA and RRF, then hands retrieval evidence back to the governed pipeline.
READ →WRITING / FIELD NOTES
Long-form thinking
Architecture notes, principles and lessons learned will live here without competing with live system state.
CONTENT SOURCECONTROL REGISTRY
Micro Brain Exists Because a Registry Cannot Know the Whole World
LightBI needs to recognize concepts that were never hardcoded, but I did not want that flexibility to become permission to guess.
READ →
Let AI Operate a Computer, but Keep the Final No on the Device
A useful remote system is not just one that can execute commands. It must know which device is targeted, who authorized it, and still let that device say no.
READ →
Teaching a Machine to Understand Data Without Letting It Guess
I want LightBI to become smarter, but the smarter it gets, the more important it becomes for the system to know when to stop and say there is not enough evidence.
READ →
Light Remote Started Because I Was About to Lose My Usual Remote
I originally needed ChatGPT to run a few commands on a VPS. A few iterations later, that temporary bridge had grown into a device system with Walls, Fleet, real PTY sessions and local policy.
READ →