RevopsML

Flowchart Diagram
Product Screenshot

Inspiration

Having worked a lot with revenue operations folks, we've seen across many orgs that product analytics alone often doesn't cut it when trying to predict customer buying behavior. Many orgs have dedicated ML and data teams that tackle the problem of combining data from disparate datasets (i.e. salesforce, amplitude, or a data warehouse) and creating models over it - a tedious and expensive process that we believe can be made faster.

What it does

RevOps ML allows rev-ops teams to train custom ML models over their revenue data. We integrate with Salesforce and Amplitude, then have an ML-based feature engineering algorithm that extracts relevant data points, cleans / filters data, then feeds it into a classification model based on a user goal (i.e. detect churn in customers).

How we built it

Revops ML uses xgboost for ML classification and sagemaker for LLM data processing on a flask python backend. Our frontend is built with Next JS.

For data, we relied on the B2B messaging app dataset (we took a 2 week segment) and reverse engineered a Salesforce dataset.

Challenges we ran into

We only ended up using about 2 weeks of data which was around 500,000 rows. Processing and storing this data efficiently proved to be harder than expected, and if we had more time we'd likely figure out some novel way to do filtering before loading data into a pandas frame for processing.

Accomplishments that we're proud of

This was the first time either of us worked with Python or traditional ML classification algorithms, but we had Chat GPT to thank for helping us figure out what we were doing while building.

What we learned

Explainability was one of our product north-stars, and we saw there's often surprising behavior trends that can go into things like churn or upsell. We also learned a lot about dealing with large datasets, and taking a more research based jupyter notebook and turning it into a more productized application.