Leo joined Partly in January 2024 and is a Data Scientist in the Classification team.
Q: What does the classifications team do?
The purpose of the classifications team is to classify the world’s parts. We get a bunch of raw data about a vehicle part, which we need to convert into complicated classifications, like part types, where the part fits into the vehicle, whether the part is a member of an assembly and more.
To give you an idea of the scale of the data we are working with, Toyota has ~2.9 million parts across ~300,000 build sheets. This is just a single OEM, to have data on 99% of vehicles in New Zealand, you would need data from 34 OEMs.
Q: Can you share more about how you do this?
While I can’t give too much away, I can share our workflow.
- We first source and research data from multiple sources including government databases, vehicle data research companies, vehicle builds and much more.
- We then have a in-house team of ex-mechanics and vehicle experts who manually convert the raw data into the classifications mentioned above. Our team classifies around 600 unique parts per day using a custom classification tool that we have developed internally. This tool allows for extreme flexibility, enabling us to make big changes very quickly according to the QA team’s needs.
- Finally, we take the raw data and train large, complex state-of-the-art deep learning models on the classifications that the QA team has manually assigned. These models use text, images, and numerical features to try to learn how to think like a mechanic. A key challenge here is managing the complexity of the input and output data. We receive large quantities of highly unstructured data and we need to turn it into something extremely structured.
Q: What do you enjoy about classifications?
I didn’t want to spend a decade researching deep tech before seeing my work in the market. Applying machine learning at Partly and seeing its immediate impact on our customers is both rare and exciting.
We also work closely with other teams at Partly.
- We work with Parts Interpretation to understand how we (Partly) represent parts, both internally and to the fitter. Understanding context around vehicle parts, like what the name of the part should be and how it interacts with other parts is key to making good classifications. It doesn’t matter if your model is 100% accurate at classifying vehicle parts if the fitter can’t understand the classifications you’re assigning.
- Work with QA to organise how we are collecting our machine learning training data and what is being collected.
- We collaborate with the Commercial team to create a feedback loop on how the classifications are perceived by the repairer. This allows us to adjust our models to align with fitter expectations.
