← All speakers

Rishabh Bhargava

Conference affiliation: Together AI · 2026

Rishabh Bhargava co-founded Refuel.ai with Nihit Desai and led it as CEO. Refuel built an open-source library and specialist language models that turned unstructured text into labeled examples for training and evaluation, or into fields that software could use. Together AI acquired Refuel in May 2025. By 2026, Rishabh was leading Together’s voice AI team, where the problem was getting several models to respond together at the pace of a conversation.

On this page

From extracting information to building data pipelines

Rishabh studied at Cambridge and earned a master’s in computer science at Stanford, where he met Desai. In 2017, he joined Primer.ai as an early machine learning engineer, working on summarization and entity extraction. His team built a system that could take a question and assemble a report from material such as news, social media and SEC filings. At Datacoral, he worked across engineering and product before leading sales engineering, helping prospective customers test whether they could move and transform data without waiting for a dedicated data engineering team.

At Refuel, Rishabh served as CEO and Desai, who had led machine learning teams working on content integrity at Meta, served as CTO. They set out to reduce the manual work of preparing datasets: assigning labels, pulling consistent fields out of documents and checking the results. At Refuel’s public launch in June 2023, the two introduced a labeling benchmark and Autolabel, an open-source Python library that let teams choose a language model, run it on their own data and measure how well it performed on their examples. Both contributed to the project. The release history credits Rishabh with documentation and release work and Desai with model support and validation work.

Deciding which labels need review

Automating the first pass still left teams with a decision: which labels could they accept, and which should a person check? Refuel tested whether models could assess their own answers by reporting a confidence score. Token-level probabilities, the model’s estimates for the pieces of text that make up a label, did a better job of separating correct labels from errors. Autolabel used that signal so teams could set a review threshold for the task at hand.

Refuel also trained its own models for recurring jobs such as classifying text, extracting named fields and deciding whether two records refer to the same entity. In May 2024 the team released Refuel LLM-2 and published weights for a smaller Llama-based companion. The team used Autolabel to evaluate the models and added nonpublic test sets to reduce the risk of testing on examples the models had already seen. In January 2025, Rishabh introduced Refuel LLM-2-mini, a 1.5-billion-parameter model, arguing that smaller specialist models could make recurring data transformations cheaper to run. When the Refuel team joined Together AI that May, Refuel LLM-2 became available on Together’s platform for serverless inference and fine-tuning.

Bringing voice models closer together

In March 2026, Rishabh co-authored the launch of Together’s co-located voice infrastructure. His example of an already optimized voice agent shows why the team focused on placement: speech recognition, language-model inference and speech synthesis can each be fast, yet the caller still waits while the software coordinating the call reaches across data centers. Run the models close to that software and the caller can hear a reply sooner, without changing any of the models.