Blog

Notes on the systems I build, with the figures produced by the code beside them.

Most of these follow one thread, six weeks of building a machine learning system a week from scratch. They are listed newest first, so if you want them in order, start with word2vec. The two deep dives at the top stand on their own.

Two architectures side by side. Late fusion runs each input through its own network before combining them; early fusion concatenates everything first and uses a single network.

Early vs Late Fusion: The Gap Closes as the Model Grows

multimodal
fusion
deep-dive

Early and late fusion are usually argued as two architectures. They are one architecture with a dial, and most of the difference is about width.

Jul 27, 2026
9 min

A scatter plot. Six pole words sit evenly around a dashed hexagon, and six corner words sit just outside it, each near the poles it is made of. The remaining two corner words, queen and boy, sit on top of each other at the centre, inside a dotted circle.

Superposition in a Toy word2vec: Three Attributes, Two Dimensions

nlp
embeddings
interpretability
deep-dive

A model can represent more features than it has dimensions. This is the geometry it uses to do it, and what it costs when the features have to share.

Jul 20, 2026
7 min

RLHF with PPO from Scratch: Four Models, GAE and LoRA

rl
rlhf
language

The reference, the policy, the reward model and the value model, and the arithmetic that turns one score for a finished summary into a gradient per token.

Jul 13, 2026
13 min

Whisper's multitask training format: a token sequence branching through previous-text prompting, a language tag, transcribe versus translate, and timestamp options.

Speech Emotion Recognition with Whisper: Classifier Head vs Vocabulary Token

audio
speech
architecture

Two designs over CREMA-D, three seeds each, and what changes when the label is a class index instead of a word the decoder has to say.

Jul 9, 2026
3 min

Fine-tuning Whisper Without Catastrophic Forgetting

audio
speech
fine-tuning

How sound becomes something a transformer can read, and what a handful of gradient steps on one clip actually does.

Jul 6, 2026
8 min

Contrastive Pre-training and a Fine-tuned Adapter: Captioning with CLIP and Qwen

multimodal
vision
language

A frozen CLIP encoder, a frozen Qwen decoder, and the 1.5M-parameter adapter between them that is the only thing trained.

Jun 29, 2026
5 min

The full encoder-decoder: a three-digit image through linear projection of flattened patches into a stack of encoder blocks, the encoder output feeding decoder blocks alongside the tokens generated so far, producing 7 7 6 9 and then a finish token.

Encoder-Decoder Transformers: Sequence to Sequence, Variable Output Length

vision
transformers

A decoder that emits digits until it emits an end token, trained by teacher forcing and scored on whether the whole sequence is right.

Jun 25, 2026
5 min

Patch Embeddings: Making an Image a Sequence

vision
transformers
attention

What attention does, why the architecture is called a transformer, and what the position embeddings are worth once the patches are in place.

Jun 22, 2026
7 min

The History of Search: From TF-IDF to PageRank to Two-Tower Models

retrieval
embeddings

Two towers over query and document, and what changes when every other document in the batch becomes a negative instead of one drawn at random.

Jun 15, 2026
8 min

Predicting Hacker News Upvotes from 872,000 Posts

multimodal
nlp

A title, a timestamp, a link and an author against 872,556 real posts: what each input is worth, and why most of the outcome is in none of them.

Jun 11, 2026
5 min

word2vec from Scratch: Turning the Meaning of Words into Numbers

nlp
embeddings

Skip-gram and CBOW with negative sampling, written from scratch and trained on text8, then scored against WordSim-353, SimLex-999 and the Google analogy set.

Jun 8, 2026
11 min
No matching items