My desktop has an AI companion now. The hard part was teaching her to say no.
Getting a local model to act was easy. Deciding what it must never decide was the work.
The Chronicles
The blog of Bhargavaram Krishnapur. Each post is about something I built: what it does, why it works that way, and what went wrong on the way.
Painting: John Martin, Ruins of an Ancient City, c. 1810–20. Cleveland Museum of Art, CC0.
Testing every finding against pure noise
Run enough tests and you will always find something. The honest move is to show what noise finds too.
My basketball model scored 0.97 AUC. It was cheating.
When a result looks too good, suspect a leak before you suspect genius.
Building an AlphaZero-style chess engine in Python
No opening book and no human games. It plays itself, and it learns from the result.
The hour of silence: a UX study of campus visitor entry
The form was annoying. The real problem was a step that nobody owned.
Local-first software
A tool you cannot run offline is a tool somebody can switch off.
Making a market signal explain itself
A reading is worth nothing if it cannot say how it was taken.
Writing a transformer from scratch
You do not understand a diagram until you have had to make one work.
Neuroevolution: networks that grow instead of learn
No gradient. No teacher. Just mutation and whatever survives.
Browser hardening, and saying what you cannot enforce
A green tick you have not earned is worse than no tick at all.
Coding war crimes: learning from deliberately terrible code
A rule you have only been told is a rule you will break under pressure.