Notebooks & Simulations

Here we prove that what we do is not a coin flip. Some of the demos are not bootstrapped and learn in your browser - this is the goal; give them a couple minutes to fit and see how they improve overtime. And some of the demos are part of the series and improve gradually so you can compare them and see the imrovement. All the demos have (?) a short help that explains the regime and other details.
source code

Interactive · Face-ID

On-device face detection, landmark lock-on, emotion classification, and identity matching in the browser. MediaPipe FaceLandmarker + FaceNet embeddings via LiteRT JS — no training, nothing leaves the page.

Simulation · Kitchen (Non-Linear Bandit)

A tree-ensemble bandit learns which items to pick up and which to skip when an item's value flips with the time of day. Watch the TS-tree arm beat the linear foil on the same features — non-linear means interaction-aware, not neural.

Simulation · Reflex Agent

Reflex Agent (Decision Tree Ensemble) - learns what the gold is and then picks it up as soon as discovers it. A dummed down version of a "robot vacuum brain". Compare it with Mixture of Bandits. Fits within a 30 seconds.

Simulation · Active Perception

A pretrained GRU DRQN collects value-graded gold on a grid it can only sense noisily along its trajectory. A diagnose-then-mask loop gate escapes oscillation basins; toggle it to watch the agent recover.

Simulation · Mixture of Bandits

A gated linear Mixture of Bandits agent learns to collect gold, in real time. Watch it improve in ~5 rounds - give it 2 minutes to see the trend.

Simulation · Gold Digger

A linear Fitted Q-Iteration agent learns to collect gold on a grid, in real time, in your browser. Watch it improve in ~5 rounds. Proves it beats random and demonstrates the limitations.

Simulation · Bandit Fisher

A UCB agent learns to fish in real time. Watch exploration collapse into exploitation as the regret curve bends.

Notebook · Fitted Q-Iteration

Python Notebook for Fitted Q-Iteration Agents.

Notebook · Bandit Recommender

Realtime Recommendation Engine.

Interactive · Gradient Descent

Building intuition for regression, derivatives, partial derivatives, and gradient descent.

Interactive · Evaluation Metrics

R², MSE/MAE, accuracy, precision/recall/F1, and ROC/AUC — five metrics, five interactive graphs. Drag the outlier, sweep the threshold, watch AUC climb.

Capstone · Breach Risk & Stock Listing

1st ML capstone project - predicting breach risk and honest analytics [A small security dataset].

Notebook · Smoke Test

Python in the browser — numpy, pandas, scikit-learn, and matplotlib running via Pyodide, no server needed.