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.
So I build tools that make neural networks less mysterious — because interpretability shouldn't stay in research papers, it belongs in production. (And in zombie games. And in mushroom generators. You'll see.)
Why I'm really doing this → the manifesto
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:
🚀 My main project is an opensource model interpretability lab - I develop it here and experiment with it on a real production app
The stack that makes a model legible in production — from watching, to diagnosing, to fixing.Watch it think (brainscope) → diagnose why it did that (causal replay + the lens) → fix it at the source with a calibrated steering vector, receipts attached (hidden-directions → hotwire-vllm). Observability first; intervention last, once the instruments have earned it.
Tip
Don't read — just do it. The core of the lab is on PyPI:
pip install hidden-directions brainscope hotwire-vllm
— the vector factory with its eval framework, the live lens server, and the production vLLM steering plugin. Everything below runs on CPU or a free Colab GPU:
brainscope --model tiny— a browser view of a model thinking (CPU is fine)make demoin steering-mechanics — real measured figures, no GPU at all- point your own OpenAI client at brainscope — watch your app's live traffic
No account, no course — install and look.
The lab runs one pre-registered research question:when does a steering vector generalize from calibration to deployment — and what do steering evals actually measure? The hypotheses were written before the data, and they're allowed to lose.
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)"])
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"]
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
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"
classDef engine fill:#1f6feb,stroke:#1158c7,color:#ffffff;
classDef exp fill:#8957e5,stroke:#6e40c9,color:#ffffff;
class bs,hd,hw engine;
class st,tm,sm,on exp;
The blue boxes are the instrument.brainscope hosts any Hugging Face model and streams its internals to the browser; hidden-directions makes the steering vectors — auto-calibrates them (Optuna, with a KL damage guard), bakes them into weights, then audits for the bake; hotwire-vllm takes those vectors to production — steering inside vLLM's CUDA graphs, per request, steered speed = vanilla vLLM. All three speak one steering spec: a vector calibrated under the lens deploys unchanged, and a misbehaving production conversation replays back under the lens.
The purple boxes are experiments run under that lens.steeropathy wires agents together through activations instead of text; in-two-minds catches an agent hesitating between tools before it commits; steering-mechanics asks how steering vectors actually work inside the model; old-news is about a system prompt losing to the conversation history that outlived it — change a rule mid-product and the old one is still in the context, still being obeyed.
Collaborators and users — not a job (see the manifesto).
If you build on LLMs and want to see inside your model, or you work on
steering / interpretability and want to compare notes — or run
SteerBench
against your own method — open an issue on any repo and say hi. The single
best thing you can do: pip install, try it, and tell me where it breaks.
🔬 Also on the bench — smaller, self-contained ways to look inside
- 🔦 tournament-watermarking — the watermark Anthropic now puts in Claude’s output, running live as a decaying newspaper: two columns from the same model, one of them watermarked, and a torch you hold to the page to find out which
- 📜 paper-remembers — Hopfield's 1982 paper, running live: rub out any part of the page and watch it rebuild itself
- 🎭 sixteen-voices — how a tiny transformer encodes writing style, through LoRA adapters and attention heads
- 👁️ show-me-your-attention — attention maps and neuron activations over your own prompt
- 💥 detektor — the collision chamber above, open source (SmolLM2 in your browser, no server)
- 🖼️ jepa-demo — I-JEPA & V-JEPA 2 hands-on, no GPU needed, with a visual deep-dive article
- 🍄 Mushroom-generator — a VAE growing mushrooms, with latent-space walks and the decoder taken apart layer by layer
- 🍎 Applepear — apples vs pears in a tiny CNN, activations and grad-CAM included
- ⚙️ Minimize_me — race TensorFlow optimizers across loss landscapes
🃏 And off the bench
- 🎸 unlived — a gamebook of the life you never lived: it writes that life as a playable story, and you win by quitting the game to go live it for real
- 🎉 promptparty — born mid-hackathon, shipped the same day: a dashboard for friends agent-coding in one room that says when to 🗣️ TALK (all agents cooking) and when to ⌨️ PROMPT (someone's agent is waiting). Claude Code hooks report automatically;
pip install promptparty - 🎨 personal-rembrandt — you can't build a personal brand, so build a personal Rembrandt: paste your bio, GPT-2 reads it in your browser, and its activations repaint his 1659 self-portrait.
- 🛏️ go-to-damn-bed — a Claude Code skill that sends you to bed like a mom sends naughty children: it saves your work into TOMORROW.md, then counts to three. It never says what happens at three
- 👑 KingOfDiamonds — the King of Diamonds game from Alice in Borderland, played by LLMs in character, recursive strategic thinking and all
- 🗨️ paralel-discordverse — your company's Discord gets a parallel universe, populated entirely by fictional colleagues
- 🌧️ coalescence — a raindrop landing on an already-wet window doesn't splash and doesn't soak in, it stops being a drop and becomes film. The thin-film equation, solved and rendered in your browser: there is no drop in the equation, the drop is only where it starts
- 🧮 least-squares-method — code archaeology: a printed Pascal listing, photographed page by page and revived on Turbo Pascal 5.5
None of it is perfect. That's kind of delightful.