Alex HallidayIndependent mind. Connected systems.
Let’s talk
ADAPTIVE SYSTEMS LABCOSSY179 / PORTFOLIO V.03

Alex Halliday / Developer & experimenter

Alex Halliday. I build systems
that learn.

Machine learning, high-performance software,
and interfaces that make the complex feel simple.

Explore selected work GitHub

A question worth testing

I test them
under pressure.

A million possible moves. An uncertain outcome.
What matters is how a system makes its next decision.

Inside the search engine

Understanding is part of the build

Then I make
their decisions
legible.

The evidence, the tradeoffs, and the things
that still need work. All open to inspection.

Explore model interpretation

One mind. Many moving parts.

Alex
Halliday.

Machine learning, systems, and product engineering.

Start a conversation
SCROLL TO EXPLORE

Selected experiments

Curiosity,
put to work.

Some systems learn to play.
Some help us see more clearly.
Every one starts with a better question.

SEARCH / INTELLIGENCEC++ · PYTHON · PYTORCH GEOMETRIC
RESEARCH ENGINE

NAGSNeuro-Adaptive Graph Search

A chess engine that changes how it searches as the position changes.

Classical search meets graph neural networks. An adaptive controller chooses where to spend its next move.

3.82Mnodes / second
perft depth 6Read report ↗
View source Repository-reported benchmark
YOUR MOVEWHITE TO PLAY

One move. Checkmate. Can you see it?

Inspect the machinery
POSITION→GRAPH EVALUATION→ADAPTIVE SEARCH→MOVE

C++17 bitboards and transposition tables handle the search. A Python service supplies graph-network evaluations, while a bandit controller chooses between alpha-beta and MCTS. The reported 3.82M figure measures move generation, not playing strength. Next question: how much does the hybrid controller improve play at a fixed time budget?

Public research implementation. The position above is a small handcrafted puzzle, not a live instance of the engine.

LEARNING / UNDER UNCERTAINTYPYTHON · RAINBOW DQN · CUDA
POLICY / 021OBSERVE

HOLD SPACE OR PRESS AND HOLD

Interactive curriculum illustration. No model runs in your browser.

Learning
when to play.

BlackJack / Reinforcement learning

A reinforcement-learning agent taught difficult decisions in stages.

Start with simple actions. Introduce harder choices only when the policy is ready. Learning has a curriculum, too.

40,000episodes in the
documented training run
View source Training report ↗
Inspect the machinery
SIMULATION→LEGAL ACTION MASK→RAINBOW DQN→CURRICULUM

The repository combines a blackjack simulation with noisy networks, distributional value learning, prioritized replay, and progressively available actions. Its curriculum report records 40,000 episodes and a negative final expected value. This is a study of learning under uncertainty, not a promise of profitable play. The simulator's existing license and authorship remain part of the project.

PRODUCT / EVERYDAY SYSTEMSREACT · TYPESCRIPT · INDEXEDDB

Life, with a
little more signal.

Life Dashboard

One local-first place for focus, habits, tasks, mood, and notes.

View source
pulse/ A LITTLE ROOM TO THINK

Make space for a good day.

An interactive study of the dashboard. Sample data stays here.

TRY MOVING A WIDGET
A moment of focus
25:00ONE THING AT A TIME
Small steps, today

Progress looks different every day.

A thought worth keeping

What if the system
made room for
being human?

A work in progress. Like the rest of us.

Route-level code splittingLocal data with DexieSortable widgetsPWA foundations
Inspect the machinery

React 19 and TypeScript provide the interface. Zustand handles application state and Dexie stores tasks, habits, sessions, and notes in IndexedDB. The source includes lazy routes, error boundaries, and PWA configuration. This portfolio includes a small interface study, not an embedded copy of the full application. Next step: test the local-data export and recovery experience with real users.

RESEARCH / EXPLAINABILITYSCIKIT-LEARN · SHAP · CALIBRATION

A prediction
should come
with context.

Heart Disease / ML pipeline

A prediction pipeline designed to show its work and its limits.

Thirteen features. A small research dataset. A careful look at what the model knows, how sure it is, and where it falls short.

View source
THE THRESHOLD QUESTIONILLUSTRATIVE DATA

Where would you draw the line?

Bar height shows the synthetic probability. Move the threshold to change which examples count as positive.

LESS LIKELYMORE LIKELY
Precision75%
Recall75%
Flagged positive12 / 24

Synthetic examples, not patient data or model results. Cyan marks actual positives; pink marks predicted positives.

Research demo. Not a medical device or diagnostic tool.

Inspect the machinery
PREPROCESS→CROSS-VALIDATE→CALIBRATE→EXPLAIN

The eight-notebook pipeline covers 270 records and 13 clinical attributes. It includes stratified validation, probability calibration, SHAP explanations, and demographic performance checks. Published README performance ranges are not reproduced here as measured results. External validation is still needed before any clinical use.

Still in orbit.

Other questions I have been following.

NBA Prediction LabSports data, model ensembles, and backtesting. Extends an existing open-source project.APPLIED MLMario RLA tutorial-led experiment in learning through play.REINFORCEMENT LEARNINGEvery experiment on GitHub

Connected disciplines

The interesting part is the intersection.

I like the space where an algorithm becomes a system, and a system becomes something someone can use.

Learning is only part of the job. Evaluation, calibration, and understandable predictions make the result useful.

Behind the systems

Hi, I’m Alex.

I follow the questions
that don’t leave me alone.

Based in St Austell, England. I build software, explore machine learning, and enjoy the kind of problem that asks you to think a little differently.

My work moves between C++, Python, and the browser. The common thread is curiosity: how does it work, where does it break, and how could it be better?

A little more about me
GOOD QUESTIONS WELCOMETHE NEXT EXPERIMENT STARTS HERE

Have a difficult
system in mind?

Engineering roles, selected freelance work,
or a collaboration worth testing.
I’d like to hear what you’re thinking.

Start a conversationalexhalliday@outlook.com
View GitHub ↗
LAB CONSOLE / 179

Welcome, curious mind.

Type help for the map. Some doors are less obvious.

NO COMMANDS LEAVE YOUR BROWSERESC TO CLOSE
ARCADE SHIFT / ROUTE THE SIGNAL

Connect the pink start to the cyan receiver. Arrow keys or the controls below move the signal. Avoid the dark cells.

Signal waiting. Find a path.