AI Engineer • Former Particle Physicist ⚛️ • Former Risk Modeler 📈
I've spent my whole career looking inside systems that would rather stay opaque — first particle collisions, then risk models, now neural networks.
The models talk to us all day. I want to be able to affect them back.
And I want to shove AI into the device and look inside it: edge AI and interpretability, together.
Why I'm really doing this → the manifesto
Three ways in:
| you want to | go to |
|---|---|
| 🔬 see inside a model: watch it think, find out why it did that, fix it | the interpretability lab |
| 🔩 run a model on the device: in a phone, or on a chip that costs less than lunch | edge AI |
| 🎈 just play with something: mostly small models doing odd jobs in a browser tab | the cabinet of drawers |
Five minutes, no account, no chip:
pip install brainscope && brainscope --model tiny # watch a model think, in your browser
git clone https://github.com/moudrkat/maxwell-on-a-chip && cd maxwell-on-a-chip
cc -O2 -o host/maxwell host/maxwell.c -lm && host/maxwell 28 # Maxwell and Boltzmann at 28 °C
My personal site is a chat with a tiny LLM running entirely in your browser, and every answer it generates renders as a real particle collision. Click the event below to fire your own question into the chamber:
Watch a model think (brainscope, probascope) → diagnose why it did that → fix it at the source (hidden-directions → hotwire-vllm). Observability first; intervention last, once the instruments have earned it.
Tip
Don't read, just do it. pip install brainscope, then brainscope --model tiny: a browser view of a model thinking. CPU is fine. No account, no course.
Click any box to open its repo.
flowchart TD
hd["🧭 <b>hidden-directions</b><br/>behavior → vector"]
bs(["🧠 <b>brainscope</b><br/>watch a model think<br/>(on your running app)"])
ps(["📱 <b>probascope</b><br/>watch a model think<br/>(inside your phone)"])
st["🕹️ <b>steeropathy</b><br/>agents talking through activations"]
tm["⚖️ <b>in-two-minds</b><br/>agent hesitating between tools"]
hw["🔥 <b>hotwire-vllm</b><br/>steering in production vLLM"]
sm["🧪 <b>steering-mechanics</b><br/>how steering actually works"]
on["📰 <b>old-news</b><br/>stale history outranking<br/>your system prompt"]
kj["🦵 <b>kneejerk</b><br/>decision models as a library<br/>(llama.cpp, no server)"]
hi["🎯 <b>hidden-intent</b><br/>a probe that reads intent<br/>before the first line"]
hd -->|"vectors"| bs
bs -->|"hosts & captures"| st
bs -->|"hosts & captures"| tm
st -.->|"steers with"| hd
hd -->|"vector + passport"| hw
bs <-.->|"same spec: lab ↔ prod"| hw
bs -->|"causal replay"| sm
hw -.->|"vector under study"| sm
bs -->|"hosts & captures"| on
bs <-.->|"the lens: server ↔ phone"| ps
kj -->|"model inside the app"| ps
hi -->|"probe"| bs
click hd "https://github.com/moudrkat/hidden-directions"
click bs "https://github.com/moudrkat/brainscope"
click st "https://github.com/moudrkat/steeropathy"
click tm "https://github.com/moudrkat/in-two-minds"
click hw "https://github.com/moudrkat/hotwire-vllm"
click sm "https://github.com/moudrkat/steering-mechanics"
click on "https://github.com/moudrkat/old-news"
click kj "https://github.com/moudrkat/kneejerk"
click ps "https://github.com/moudrkat/probascope"
click hi "https://github.com/moudrkat/hidden-intent"
classDef engine fill:#1f6feb,stroke:#1158c7,color:#ffffff;
classDef exp fill:#8957e5,stroke:#6e40c9,color:#ffffff;
class bs,ps,hd,hw,kj engine;
class st,tm,sm,on,hi exp;
Research that only runs behind a server isn't finished. These take it the last mile, natively. No cloud, no account, nothing phones home.
- 🦵 kneejerk — decision models as a library on llama.cpp: no server, one forward pass, built for a phone app.
- 📱 probascope — an Android app: a decision model runs in the phone, and you drag over its layers to see where the answer formed.
- 🧭 decide-me — point your phone at the world and a decision model inside it tells you what to do next. I said I would not build this.
An ESP32-S3: 16 MB of flash, a few dollars, no operating system. Models of 3–4 million parameters, trained from scratch, running on it.
- ✏️ esp32-ai — you type a sentence, a 3M-parameter model inside the chip writes a small program for its own screen, and the chip runs it. Built on slvDev/esp32-ai.
- 🌡️ maxwell-on-a-chip — a tiny model of Maxwell and Boltzmann, sampled with their own distribution, at the temperature of the room. Maxwell's demon is on the BOOT button.
Small models on small devices are the ones you can actually open up: probascope is the lens in a pocket, and hotwire-vllm is the same last-mile move for steering, in production vLLM.
Collaborators and users. Build on LLMs and want to see inside your model? pip install, try it, and
open an issue where it breaks.
What would you use it for? → tell-me: my guesses, and questions for you
Smaller things, and things that got out of hand. Pull any handle.
━━━ 🌐 In a browser tab — what a small model can do with no server behind it
|
━━━ 🔬 Also on the bench — smaller, self-contained ways to look inside
|
━━━ 🃏 And off the bench — everything else that got out of hand
|
None of it is perfect. That's kind of delightful.