Blog Drafts

The classic 2D Snake game is already a decent RL benchmark - sparse rewards, self-collision, and a growing body that changes the game state. But 2D Snake is one

CartPole is a pendulum balancing problem. The AI keeps it upright by applying force.

Most RL benchmarks don't have a known optimal policy. You train, you measure win rates, but you never know how far off you are from truly optimal play.

Breakout is where DeepMind's DQN first shocked the world in 2015. Their agent learned to play Atari games from raw pixels alone, eventually discovering the "tun

Tic-Tac-Toe has 255,168 possible games. Connect 4 has 4,531,985,219,092. That's 18 million times more states.

Flappy Bird is a meme. Two pipes with a gap. Tap to survive. A seven-year-old can figure it out in ten minutes.

A rat in a maze learns by trial and error. Put it in a new maze and it explores again. Our DQN agent does the same - except it trains on thousands of random maz

Pong is the original video game and arguably the cleanest testbed for learning from interaction. Two paddles, one ball, one rule: don't let the ball past your p

CartPole solved in 10 generations. Tic-Tac-Toe achieved perfect play. So I built a 2D rocket lander and expected DQN to crack it in an afternoon.

The multi-armed bandit problem is where machine learning meets Vegas. You've got a row of slot machines, each with an unknown payout probability. You've got 1,0

Tic-Tac-Toe is a solved game. Give anyone a few minutes with a strategy guide and they'll never lose again. So why train an AI on it?

Dropping a package from a moving plane sounds simple. The physics is taught in first-year university: time to fall = sqrt(2h/g), horizontal distance = v t. Tri

DQN learns by computing gradients. It adjusts weights in the direction that reduces loss. It requires a differentiable loss function, a carefully tuned learning

Mountain Car is one of the simplest RL benchmarks: a car in a valley needs to reach a flag at the top of a hill. Two actions: push left, push right. The car isn

I built a self-play training loop where an agent starts knowing absolutely nothing about tic-tac-toe, plays itself thousands of times, and emerges with a perfec

I implemented multi-head self-attention from scratch three times before I actually understood what was happening. Then I built a visualisation showing which tok

I've always been fascinated by murmuration - that hypnotic dance thousands of starlings do at dusk, moving like a single organism. Last month I implemented Crai

Draw a rubber band around a scatter of points. Tighten it. The polygon that forms is the convex hull - the smallest convex shape that contains all your points.

I spent two weeks trying to understand diffusion models by reading papers and getting nowhere. Then I built a minimal DDPM (Denoising Diffusion Probabilistic Mo

I stumbled into DuckDB last month while frustrated with pandas groupby being slow on a 2GB CSV, and it's genuinely changed how I explore data. The hook: it's a

I implemented Jos Stam's "Real-Time Fluid Dynamics for Games" paper and watched ink swirl realistically through water without solving a single differential equa

Procedural terrain is magic. You start with random values, apply some fractal logic, and out comes mountains, valleys, and coastlines. No artist required.

I've been obsessed with procedural generation lately, and L-systems are the cleanest way I've found to generate plant-like structures. Feed in a rewriting rule,

I trained a PPO agent to play Carcassonne using self-play, then analysed 50 games of the trained model to see what it actually learned. The results are interest

I've been using pandas daily for six years. Last month I hit a wall with a 2GB dataset taking 45 seconds to pivot, so I tried Polars. It ran in 2 seconds. But I

I built a small desktop monitoring tool twice - once in Electron, once in Tauri. This isn't a theoretical comparison. I hit the same walls both times, and they'

Imagine you scatter points across a map. Every location on that map belongs to the nearest point. Colour each region by which point owns it. That's a Voronoi di

I built a small data explorer last month that runs entirely in the browser - no backend server, no API calls, just Python running in a tab via WebAssembly. It's

3D plotting is one of those things that looks impressive in a paper but can be genuinely hard to get right. I've spent the last year generating 3D plots for sci

Point clouds are just lists of 3D coordinates. A LiDAR scanner spits out millions of them per second. The challenge: given a chaotic blob of 3D points, figure o

I needed to know: what's the largest circle that fits entirely within Australia's borders? Not a theoretical exercise - I was curious about the actual geometry

Blender is a 3D modelling powerhouse, and it has a Python API (bpy). I spent a week learning it, and it's equal parts brilliant and frustrating. You can script

Built a real-time dashboard with Bokeh last week. It's the kind of project that looks simple until you're debugging the layout for an hour and wondering why the

I've spent the last five years fighting with these libraries, and I finally have opinions worth sharing.

A car park at 80% capacity feels full. At 95%, it feels chaotic. But the maths doesn't match the feeling. I built a simulation to understand why.

I've built a few React Native apps over the years, and every time I'd approach them the same way: set up bare React Native, fight Gradle, compile for two hours,

I spent a weekend fine-tuning SDXL to generate images of a specific style I wanted, and it was weirdly satisfying and deeply frustrating in equal measure. If yo

I wanted to convert photos into single continuous line drawings - the kind of minimalist art where one pen never lifts off the paper. The technical challenge: g

I spent an afternoon turning a flat DEM (Digital Elevation Model) into something that actually looks like a map. You know those squiggly lines on topographic ma

An isochrone is the answer to: "How far can I travel in 15 minutes from here?" It's a boundary line (or polygon) that encloses all reachable points. They're use

I've been obsessed with emergence lately - how simple rules create complex behaviour. The canonical example is flocking: three rules about separation, alignment

I wanted to build something simple: overlay different lakes on top of each other at the same scale so you can actually compare their sizes. The catch is that la

I implemented Levenshtein distance from scratch for a spell checker project, and it's genuinely satisfying how much computer science is contained in a simple gr

Mazes are a perfect excuse to implement classic algorithms. You generate them with recursive backtracking, then solve them with A and BFS. The fun part is anima

NEAT (NeuroEvolution of Augmenting Topologies) is wild. Instead of training a fixed neural network, you evolve the network structure itself. Add nodes, add conn

Pandas is great until you call apply() on a large DataFrame and watch your CPU spin at 25% while the rest of your machine sits idle. Then you remember that Pyth

I built a car parking simulator in Pygame and trained a neural network to solve it. For the first 500 steps, the car did something unhelpful (usually drove into

Built an animated 3D scatter plot the other day that showed a trajectory evolving through time. Sounds simple, right? It's not, but Plotly makes it less painful

PPO (Proximal Policy Optimisation) is the workhorse of modern RL. It's stable, sample-efficient, and doesn't require as much hyperparameter tuning as DQN. I tra

Kafka is the industry standard for event streaming, and it's also kind of a pain. It's written in Scala, it's slow to start, it requires ZooKeeper (which is its

I've been building full-stack apps in JavaScript for years, and the pattern is always the same: separate frontend repo, separate backend repo, maybe share some

I spent a weekend simulating river erosion on an HTML canvas. The result looks like a real river snaking across a landscape. No pre-rendered assets, just physic

I implemented three routing algorithms on a real road network and timed them. The results were... not what I expected. A didn't always win, Dijkstra was slower

I built a cellular automaton where Scissors, Paper, and Rock are actual populations fighting for space on a grid. Each agent spreads by converting its neighbour

I built an evolution simulator from scratch. Creatures that eat, reproduce, mutate, and get eaten. No fitness function we designed - just let the rules loose an

Data lives everywhere now. DynamoDB in production, MongoDB in legacy systems, Firestore in serverless projects. At some point you need all of it in one place -

Generative design for spatial layouts is the fun intersection of algorithms and geometry. The idea: define rules and constraints, let an algorithm iterate throu

I built a terrain generator that felt like magic. You give it a set of tile rules, press go, and it fills a grid with valid combinations. It's called Wave Funct

I looked at xTuring and thought "this looks suspiciously easy - fine-tuning an LLM in 50 lines of Python instead of 500?" So I spent a week actually using it on

I've been building a lot of tools lately and reaching for custom integrations every time something needs to talk to something else. But I realized I was being d