Locala builds a GenAI virtual assistant on Databricks that explains ad campaign data in natural language

Overview

Locala, an omnichannel advertising platform, fine-tuned a Llama 2 model on Databricks using MLflow, Lakeflow Jobs and Model Serving to build 'Insights AI,' a virtual assistant that lets planners and clients ask natural-language questions and receive explanations of advertising campaign performance charts and figures. A single data scientist was able to build the tool, which is now used in client-facing sales presentations, and Locala plans to evolve it into a RAG system for broader querying.

The challenge

Locala's small five-person data team faced significant challenges developing, training and deploying machine learning models, and users needed extensive expertise in Locala's massive internal platform to derive insights from advertising data, a lengthy and demanding process without a virtual assistant.

The solution

Locala used Databricks tools to streamline the ML lifecycle, fine-tuning a Llama 2 model for cost-effectiveness, using MLflow for version control and experimentation tracking, Databricks Lakeflow Jobs to orchestrate and automate ML and ETL pipeline scheduling with optimized machine selection, Databricks Model Serving to benchmark model versions, and Unity Catalog to manage access rights and governance, letting one person build the Insights AI virtual assistant.

Reported business value

Just one person was able to create the GenAI solution, significantly reducing the complexity and time required to interpret campaign data; Lakeflow Jobs lowered operational costs while freeing up time for model training, and the tool has helped Locala generate briefs faster and more accurately, playing a role in sales presented by their CEO at major industry conferences.