Projects

Personal tool

Reading River

Reading River is a way to lower the pressure to read everything. It is for people who save more articles than they can read.

The idea comes from Oliver Burkeman: a reading list should be more like a river than a bucket. You wade in, choose something that interests you, and let the rest pass.

How it works

Save a link directly or use the Chrome or Firefox extension. Each day, 16 articles are drawn from your collection. They float past as paper boats in a three-dimensional riverside scene. You can walk along the bank, look at their titles, and open one.

Opening an article moves it to the eddies, a pool that holds up to five pieces you have started reading. You can return to them later or send them back to the river. Finish an article and a golden tree grows on the bank. The trees stay where they were planted; you can walk up to one to reread its article.

The river borrows details from paintings. Monet's Japanese bridge crosses the water, and Turner's Fighting Temeraireis towed downstream. There is also a daily email containing an article from your saved collection.

The philosophy

Something can seem urgent when you first come across it. Leave a link for a few days and you may no longer want to read it. The river leaves room to choose what to read when you return, without treating every saved link as unfinished work.

You can try the guest river without an account. Each visit shows 15 articles chosen from a fixed pool of 30. To start saving your own links, join the beta waiting list.

Explore the guest river

Join the beta waiting list · How Reading River works

Machine learning capstone

Exploring Bayesian Optimisers

Bayesian optimisation and Gaussian Processes.

From 2025 to 2026, I completed a Machine Learning course at Imperial College in order to develop my ML skills. One of the required outputs was a black-box optimisation (BBO) challenge, which involved optimising eight functions. Each function took a continuous input vector and returned a single output score. We had no information on the functions except for a series of previous inputs and output.

A classic problem in both machine learning and life is the explore-exploit trade-off. Is it worth listening to a new album that you might enjoy, or relistening to an old album that you know you love? This project offered a gamified way of thinking about that trade-off. Should we explore new areas or exploit areas that already seemed promising?

There are plenty of scenarios where these decisions matter in the real world. Let's say we could send ten robots to Mars because we want to drill and collect a sample. Where should we land these robots so that our last robot gives us the highest possible yield of our sample? To give a more concrete example, choosing the hyperparameter settings of a neural network can be modelled in this way.

Play the Game.

If you're interested in the technical side, you can read the notes below or you can see the entire code on my GitHub. But to make this more fun, I also built a small browser game where you too can try to beat a Bayesian Optimiser.

In the game, the player and a Gaussian-process optimiser search the same hidden two-dimensional landscape. The player chooses points manually. The optimiser chooses points using different strategies. At the end, the shape of the landscape is revealed.

This demo is simpler than what a real project in this space might involve. Firstly, it's two-dimensional, so you can actually see the landscape - everything is harder in more dimensions! Secondly, it uses a lightweight Gaussian process rather than a full BoTorch workflow. But the game is just to give you a flavour of the project.

Play the Bayesian optimisation gameBayesian optimisation game preview

Technical note

The model uses a Gaussian process with an additive linear plus Matérn kernel, and output transformations such as Box-Cox, sign-flipped Box-Cox, or Yeo-Johnson where useful. The functions were small-data, continuous, and expensive to query, so a Gaussian process was a natural starting point.

Once the surrogate is fitted, the pipeline generates a mixed pool of candidate points:

  • Global Sobol samples, designed to preserve broad coverage of the search space.
  • Trust-region candidates, concentrating search around the current most promising area.
  • Elite-region candidates, which look around several of the best observed points, preserving promising alternatives.

Each candidate is then scored using several acquisition-style signals, including expected improvement (EI), log expected improvement (logEI), probability of improvement (PI), upper confidence bound set with a range of betas (UCB), Thompson-style scores, posterior uncertainty, and novelty.

Go to the Model Card