1. Technology overview
RunRL is a San Francisco-based company, as stated on its website, that provides reinforcement learning as a service. It addresses the problem that generic language-model agents need long instructions and are often too unreliable to deploy for a specific job.
Customers supply a reward function that scores good and bad outputs, and the platform runs reinforcement-learning fine-tuning to produce a specialist model. The launch post describes a chemistry example in which a much smaller model beat frontier models on a drug-candidate binding score, and the site lists models trained for interpretability, formal verification, drug design and automated red-teaming.
2. Highlighted features & specifications
- Reward-function training: customers define what counts as good output and the platform optimises the model against it.
- RL fine-tuning: applies the class of algorithms used for recent reasoning models to a customer's own use case.
- Agent reliability: reinforces rule-following, information extraction and document navigation for existing agents.
- Specialist models: reported work covers interpretability, formal verification, drug design and automated red-teaming.
- Managed service: the team runs the training platform and can be reached by email or a booked call.
3. Potential application
A Vietnamese fintech could train a document-extraction model against a reward based on field accuracy. A research lab could optimise a molecule-generation model against a scoring function. A software company could improve the rule-following of its customer-support agent.
4. Benefits
Teams can turn a general model into a specialist for one task without writing long prompts. A custom reward function ties training directly to the business metric. Smaller specialist models can cost less to run than frontier models. The platform removes the need to build RL training infrastructure in-house.
5. Market trends & opportunities
Enterprises moving language-model agents into production need higher reliability than prompting alone gives. Reinforcement learning on task-specific rewards is one route to specialist models. In Vietnam and Southeast Asia, teams building domain-specific AI for finance, manufacturing or research face the same gap.
6. Other information
Founders Derik and Andrew. Backed by Y Combinator (Spring 2025 batch). Site states it is made in San Francisco and lists Promontory as a backer.
One platform connecting startups, corporations, research institutions, regions and investors











Playbook & Handbook · February 6, 2026
Winning Accelerator Playbook 2026 by BambuUP
RETAIL E-COMMERCE: AI AND REAL-TIME DATA ARE TRANSFORMING HOW RETAIL BUSINESSES OPERATE








































