Ronan McGovern is the founder and chief executive of Trelis Research, where he develops practical approaches to adapting, deploying, and evaluating AI models. Before working on browser-operating agents, personalized speech, and computationally efficient reasoning, he co-founded and led Sandymount Technologies, an MIT spinout that commercialized membrane-filtration research.
McGovern studied mechanical engineering at University College Dublin and earned a master’s degree and doctorate at MIT, where he also conducted postdoctoral research on membrane desalination, reverse osmosis, and water-treatment economics. At Sandymount, he developed a filtration system that concentrated beer for more efficient transportation while preserving important flavor characteristics. Alfa Laval subsequently acquired the company and its patented Revos technology. His writing on paid customer trials distills a lesson from that experience: customers willing to pay provide a stronger signal of genuine demand than users accepting free access.
Training browser agents from successful reasoning traces. McGovern connects Qwen models to Playwright through the Model Context Protocol, records successful tool calls and conversational turns, and converts multistep interactions into training examples for smaller models. His browser-agent workflow preserves intermediate decisions instead of treating successful navigation as a single response.
Curated supervision before reinforcement learning. Carefully selected demonstrations can establish useful behavior before reward-based training begins; otherwise, a model may reach correct outcomes too rarely to learn efficiently. McGovern emphasizes held-out evaluations, realistic multistep tasks, and larger datasets over isolated successful demonstrations.
Personalized speech through model adaptation. His speech fine-tuning workflow prepares single-speaker recordings with Whisper, corrects transcripts, and adapts Sesame’s CSM-1B using low-rank adapters. The model represents sound through hierarchical audio tokens, with a primary transformer and secondary decoder handling different levels of detail. McGovern distinguishes sample-conditioned voice cloning from fine-tuning, which changes model parameters, and demonstrates how combining both can improve resemblance to a target speaker.
A browser agent’s useful training data includes its tools, reasoning, actions and observations. Ronan McGovern follows that record from collection through Qwen3 adapter training and back to inference.
What should a browser agent learn from a successful run?
Turn a single-speaker recording into paired audio and text, adapt CSM-1B with LoRA, and compare what reference audio and fine-tuning each contribute to the generated voice.
▶ Watch ↗