Delphina: Building a Magical ML Developer Experience with AI

By: Tony Liu & Madison Hawkinson

In 2015, Jeremy Hermann built the first state-of-the-art machine learning platform, Michelangelo, that enabled hundreds of Uber data scientists, including co-founder Duncan Gilchrist’s teams, to deploy and experiment across thousands of models. After seeing its impact, Jeremy co-founded the first machine learning feature store, Tecton. 

After leaving Uber, Duncan felt the ongoing pain points of manual machine learning workflows and scattered point solutions while leading data science at Gopuff. Data science workflows remain highly labor intensive and there simply wasn’t enough machine learning or data science talent to fill all the gaps.

Their conclusion: despite the emergence of the MLOps ecosystem, the process of machine learning engineering was still manual, fragmented, and siloed. And that was getting in the way of machine learning’s impact on businesses in areas as varied as fraud models to dynamic pricing to accurate forecasting. 

Jeremy and Duncan had no doubt that there must be a better experience than the weeks or even months currently spent by teams developing strong ML models. They wanted to create a co-pilot for data science that accomplishes two things:

• Create a magical developer experience for machine learning
• Enable machine learning teams to deploy models at scale without engineering drag

Delphina uses large language models (LLMs) to supercharge data exploration and modeling by optimizing feature engineering and modeling in one place. The weeks and months data science teams spend manually crafting features and tuning models goes down to hours. Likewise the weeks of engineering work to get models working into production now goes down to minutes. This enables machine learning teams to focus on the real work – translating business problems into a model. This creates an enormous unlock for AI in critical business processes.

With Delphina, machine learning teams don’t have to hand off the model to engineers to put it into production. The model is ready to be deployed in production along with the relevant data pipelines used to generate features for the model. Once in production, Delphina will automatically measure performance and share insights. Updating the model will no longer be a multi-month project but one that takes place continuously. 

As former data scientists, we understood the product vision right away and its potential to unlock the productivity of machine learning teams to create business value. What Jeremy and Duncan observed in machine learning teams was exactly what we experienced first-hand at Meta and Databricks. Their expertise and insights make them uniquely suited to building a product that feels as magical to machine learning workflows as LLMs feel to the world in generating content today. This is why we are thrilled to continue our close partnership with Jeremy and Duncan and co-lead the seed investment in Delphina with our friends at Radical Ventures.