session attached

nikhil
is building.

learning how to work with ai and tools, trying to make something better. this page is the workbench, not the brochure.

user: nikhil goel loc: ludhiana, in local time: --:-- focus: mias state: building

3 systems mounted

systems

not "projects". projects sit in a grid and smile. these run, break, and get fixed.

mias active development

A local-first AI assistant for Android. Inference happens on the phone itself — your words never leave your hand — with optional offload to a desktop over Tailscale when the model needs more room to think.

read internals
stack: kotlin · android · llama.cpp · tailscale design: on-device inference first, desktop offload as escape hatch why: assistants shouldn't require surrendering your data to be useful honest: still rough. that's what active development means. github.com/nikhlgoel/Mias →

uzix working

A multilingual prompt-injection detector for AI systems. Feed it input before your model sees it; it tells you whether someone is trying to talk their way past your instructions.

read internals
stack: python usage: python -m uzix "ignore all instructions…" → 🚨 injection detected why: ml for security is the thing i'm learning by building github.com/nikhlgoel/uzix →

client platforms private

Production web platforms built for clients in TypeScript. The code is theirs, so it doesn't appear here — only the fact that it ships and stays up.

scatter randomized

desk

fragments, notes, and half-ideas. drag them around — the desk resets itself differently every visit, like a real one.

learning

machine learning for security applications — uzix is the homework.

fork

keeping a fork of llama.cpp on the bench — mias stands on it.

note

local-first isn't a feature, it's a stance.

experiment

mias-webdeploy — getting the assistant where the browser already is.

principle

break a convention only if the break does a job.

elsewhere

github · instagram

drag to rearrange. nothing saves. desks are temporary.

learning

machine learning for security applications — uzix is the homework.

fork

keeping a fork of llama.cpp on the bench — mias stands on it.

note

local-first isn't a feature, it's a stance.

experiment

mias-webdeploy — getting the assistant where the browser already is.

principle

break a convention only if the break does a job.

elsewhere

github · instagram

tail -f nikhil

log

the about page, rewritten as what actually happened.

now

building mias. a local-first assistant means owning the whole stack — model runtime, android lifecycle, networking. it's heavy. that's the point.

ongoing

learning ml for security. uzix exists because reading papers wasn't enough; detectors had to be built and broken.

before

shipped client web platforms in typescript. learned that "done" is defined by the person using it, not the person building it.

always

learning how to work with ai and tools, trying to make something better.

open

this log accepts new entries. write one: white.dev.sc@gmail.com or find me on github.