Apollo.io LLaMA Fine-Tuning

At Apollo.io, Pranav worked on the GTM AI Assistant, a product that let users apply filters and generate responses through a conversational interface to control the Apollo UI for outbound lead generation. The use case was highly custom: the model needed to produce structured filter outputs from natural language queries in a way that general-purpose models could not reliably do.

He led a fine-tuning POC on a LLaMA model, starting with in-house dataset construction. The team built a labeled dataset of query-to-filter pairs, defining what a correct output looked like for each input. He then fine-tuned the model on this dataset and validated performance against a golden dataset, where a pass meant the model produced the correct outputs for a given query.

After fine-tuning, the model reliably produced correct outputs on the golden dataset, demonstrating that the custom use case could be addressed through fine-tuning rather than prompt engineering alone.