My First Neural Network Hasn't Learned Anything Yet
· 3 min · AI
49 lines of PyTorch, an MNIST model that can't recognise a single digit, and why I'm weirdly happy about it.
Blog
· 3 min · AI
49 lines of PyTorch, an MNIST model that can't recognise a single digit, and why I'm weirdly happy about it.
· 4 min · Claude Code
I kept losing good ideas while Claude was busy. So I built claude-message-queue, a small plugin that saves the next thing to do and runs it only after the current work is finished.
· 3 min · AI
I have spent a lot of time shipping models to phones without really understanding how they are trained. So I started learning, from the very beginning.
· 4 min · Firebase
I built a little app for making interactive greeting cards. The most interesting decision was that there are no accounts at all. The share link is the only key.
· 3 min · AI
Before a speech model can say a sentence, something has to decide how that sentence should be said. That part is called text normalization, and it's full of tiny traps.
· 3 min · On-device AI
Wake word, speech to text, a tiny LLM, and text to speech, all running on a 2018 Raspberry Pi with 1 GB of RAM. It's slow. It also works.
· 3 min · React Native
On-device AI usually still needs the internet once, to download the model. I worked on removing that last dependency for our React Native and Swift SDKs.
· 4 min · On-device AI
No server, no API key, no audio streaming from the cloud. Just a 25 MB model, ONNX Runtime Web, and a lot of small fights with WebAssembly.
· 4 min · sdk-engineering
The hardest part of SDK engineering isn't writing the SDK. It's keeping three SDKs feeling like one product while everything underneath them is different.
· 5 min · gemma
On-device AI used to feel like a demo. With models like Gemma 4, it’s becoming a real product surface. And Apple’s slower approach might actually be the right one.
· 2 min · React Native
React Native is not perfect, but it still gives me the fastest path from an idea in my head to a real thing in my hand.
· 2 min · AI
I keep coming back to the same idea: the best AI features should feel less like a demo and more like the app quietly caring about the user.
· 2 min · Writing
I wanted a place that feels like a lab notebook: small enough to maintain, personal enough to be honest, and structured enough to be useful later.