Meet MARVIN, the Robotics/AI platform for lab testing!

Inspiration

BPR: "What is my purpose?" RS: "You pass butter". BPR: "Oh my God..."

What it does

MARVIN combines our own robotics and AI software to automate commercial microscope tests (a $50B labor market).

MARVIN is highly accessible, making it simple to digitise microscope slide samples, and rapidly train-deploy AI models.

We envision tens of thousands of MARVINs across the world doing testing for environment, agriculture, resources, medicine, and research.

Our smartest people shouldn't have to be doing this manual work! Let's put them to work where the real science is done...

How we built it

To build an AI model for a particular analysis application we scanned new samples with the MARVIN platform and generated a DL model with PyTorch, Image Annotation Tooling, Google Compute, and Jupyter Notebooks.

Challenges we ran into

GIGO. Early training used segmented annotations output from an incomplete Computer Vision approach to solving the analysis problem. This led to very bad performance of the DL model which would not learn effectively. Instead we rapidly created new annotations manually which led to a great training curve and a model that performed well for our test data.

Zone Resource Error. Google Cloud Compute is running out of resources! Training was slower as we had to settle with Tesla GPU rather than A100's.

Accomplishments that we're proud of

Better than traditional CV approaches in one night of hacking! Leaning on the expertise of Phil from the AV industry to build and train ML models quickly!

Promising proof of concept that will extend it into commercial development.

What's next for MARVIN

We're going to continue to extend our dataset annotations with more sample scans, construct a better class structure, and further train our DL models. We'll also integrate the process built into a commercial process that customers of MARVIN can use to train their own models for other microscope tests.