TrainLoop

TrainLoop is a San Francisco-based post-training research and product lab that builds specialized language and multimodal models for long-horizon tasks. It works with pharmaceutical, biotech, logistics and banking enterprises to convert domain expertise and proprietary customer data into custom models, pairing a research team studying long-horizon post-training methods with a deployment team that works alongside customer subject-matter experts.

The company's stated model areas are biological reasoning models for pharmaceutical and biotech customers (aimed at diagnosis and treatment-response work in drug development), reliable agents for high-risk industries such as financial services and logistics where decision accuracy in mission-critical workflows matters, and purpose-built multimodal reasoning systems for complex image understanding and document abstraction. Its four stated research directions are life sciences, continual training (methods that avoid catastrophic forgetting), information theory (capacity-aware objectives for stable, interpretable reasoning), and evaluation and interpretability.

At its Y Combinator launch the company positioned itself as "Reasoning Fine-Tuning": a platform making reinforcement learning-based fine-tuning accessible to developers, structured as data curation via a lightweight SDK that gathers training signals from production usage, training of a reward model that teaches an LLM the preferred outputs, and deployment of the resulting model through standard APIs. Third-party directory coverage describes the same workflow, including data collection, custom model deployment and data security.

Founding story

TrainLoop was founded in 2025 by Jackson Stokes (Founder/CEO) and Mason Pierce (Founder/CTO). Stokes previously worked on performance for Google's Gemini models and AlphaFold, with work described as supporting AI search summaries and systems including Waymo; Pierce led engineering at Second (YC W23), where he worked on large-scale enterprise codebase migrations and retrieval-augmented generation systems and encountered the difficulty of fine-tuning off-the-shelf models. The founders framed the company around the gap between internal model-training tooling at large labs such as Google and OpenAI and what is available to developers deploying models in production.

Business model

B2B. TrainLoop engages enterprises in a structured research-to-production collaboration: jointly defining research objectives from the customer's proprietary data and technical strengths, advancing models, and sustaining their performance in production. Earlier positioning centered on a self-serve developer platform with an SDK, managed training and API-based inference.

Traction

The company reports partnerships with NollaMD (differential diagnosis in visual medicine), Mercor and Pathos, and states that its models frequently achieve state-of-the-art or Pareto-optimal performance on their target tasks. Team size is listed as 6 in the Y Combinator profile.

Latest developments

TrainLoop's public positioning has shifted from a developer self-serve reasoning fine-tuning platform toward an enterprise post-training research and product lab, with recent partnership write-ups covering NollaMD, Mercor and Pathos and research notes on one-step model training with GRPO and the low-rank nature of learning GSM8K.