Writing · since 2026

Notes on what I’m learning.

From pure noise to emoji faces: building a tiny diffusion model

I trained a small DDPM on a few hundred face emojis and saved a sample grid every 50 epochs. The progression from random pixel mush to recognizably composed novel emojis is the most satisfying training video I have produced.

AI / ML · Deep Learning · Diffusion Models · Generative Models · PyTorch

Mar 25, 2026

12 min read

Building self-attention from for-loops to debug a transformer

A tiny GPT was failing at squaring numbers. To understand why, I rewrote attention in plain Python, looked at the heatmaps, and surgically boosted the right scores until the model stopped getting things wrong.

AI / ML · Deep Learning · Transformers · Attention · Mechanistic Interpretability

Mar 10, 2026

10 min read

Building an autonomous customer support agent worth shipping

An LLM that can email customers, query the database, and quote policy. The interesting part is not the language model. It is the boring scaffolding around it that decides what the agent is allowed to do, when, and to whom.

AI / ML · Agentic AI · LLMs · Tool Use · Safety

Mar 4, 2026

24 min read

Compressing emojis into 128 numbers, then adding glasses to one

A small convolutional autoencoder, 600 face emojis, and a question that has been bugging me for years: does vector arithmetic on a latent space actually work the way the papers say?

AI / ML · Deep Learning · Autoencoders · Computer Vision · PyTorch

Feb 21, 2026

11 min read

A tiny neural network and four experiments I ran on it

Two layers, four 2D toy datasets, a stack of loss functions, and a NumPy reimplementation. Here is what changed in my head about backprop.

AI / ML · Deep Learning · Backpropagation · PyTorch · NumPy

Feb 1, 2026

12 min read