Riaz Lab Let’s talk
An independent AI labTest in progress

Look over
my shoulder.

Local models, frontier models, and whatever is on the screens this week.

Seen from behind at a standing desk, facing three screens: code, object detection, segmentation masks and a 3D point cloud On the bench The notebook
Test it. Measure it. Write it down.What happens at this desk

The face isn’t
the interesting part.

The screens are. Open-weight models under my desk, frontier models through their APIs. Different tests, different tools.

This is my public notebook: the apps, the data, the method, and the part that failed.

Independent lab.
Tests only.
The back of a head, glasses just visible, in front of the screens
RiazTester. Nothing else.
Background
MIT, 2023Digital transformation
Approach
Model agnosticLocal to frontier
Principle
Measure firstThen write it down

The elements
of the lab.

Models, maps, motion and the systems between them. A few of the tools behind the tests.

Models

Claude

Frontier models through the API, alongside local open-weight models.

Used forAPI tests and coding agents

Select an element to see where it fits.

Things I’ve
been testing.

Four tests on the go. Some turn into useful tools. Others show me what breaks.

Mapping

Digital twin of an island

MapLibre · OpenStreetMap · LiDAR · SAM2

Recreating an entire island in 3D — terrain, buildings, neighbourhoods — and wiring an engine onto it that answers questions you ask of a territory.

Infrastructure

Compute at home

3 × DGX Spark · RTX 5090 · vLLM · ComfyUI

Three DGX Sparks and an RTX 5090 running open-weight models at home, with no usage ceiling and no data leaving the house.

Embeddings

The YC Atlas

fastembed · 384d · t-SNE · KMeans

Every public Y Combinator company vectorised on a CPU at home, to ask what an accelerator's portfolio actually looks like from the inside. Open it →

Game

Tubyrinth

Unity 6 · mobile

A 3D pipe maze you have to memorise. Native prototype first, to see what a game designed with agents actually plays like.

Tests in progress. Published results appear in the notebook below.

Always
learning.

The tools change.
The curiosity stays.

  1. 2023

    MIT

    Training in digital transformation.

  2. Today

    Independent AI testing

    Testing models, maps, games and compute at home.

  3. Ongoing

    The open notebook

    Publishing the method, the data and the things that failed.

    Read the latest analysis

What the
numbers said.

The analysis, the evidence,
and a way to check it yourself.

Proof,
in public.

Small milestones.
Something you can actually explore.

Let’s test
something.

hello@riazlab.com

Ideas, questions, things to test.
One address. Straight to me.