Key Takeaways
- • Trajectory, a one-year-old startup founded by three Stanford classmates, raised $45 million in a Series A round at a $300 million valuation, led by Sequoia with participation from Nvidia, Bessemer Venture Partners, and Bain Capital Ventures.
- • The 18-person company is building a platform that lets businesses fine-tune open-source AI models through continual learning, using customer feedback to retrain models without requiring internal research teams or forward-deployed engineers.
- • Trajectory's customers include Harvey, Rogo, Decagon, Mercor, and Nvidia, with Harvey and Rogo using the platform to adapt models for legal and financial analysis tasks.
- • The company is now making its post-training platform publicly available, positioning itself against Thinking Machines Lab's Tinker, with Applied Compute (valued at $3.25 billion last month) and Prime Intellect also cited as competitors.
- • Co-founders with prior experience at Google DeepMind, Apple, and Windsurf-related acqui-hires argue that human involvement in retraining makes their approach safer than autonomous recursive self-improvement.
Summary
Trajectory has raised $45 million in a Series A funding round at a $300 million valuation, according to TechCrunch, led by Sequoia with participation from Nvidia, Bessemer Venture Partners, and Bain Capital Ventures. The round comes roughly one year after the startup was founded by three Stanford classmates with prior experience at Google DeepMind and Apple. The company has 18 employees.
The startup addresses a core limitation of current AI models, which do not learn from their mistakes without further training. Businesses adapting general models for specialized fields such as law and finance typically fine-tune them, a process that often requires internal research teams or costly consultants known as forward-deployed engineers. Trajectory aims to automate this work through continual learning. Companies connect the platform to their applications and start with an open-source model. The platform collects feedback on the model's performance and uses it to retrain the model, improving its handling of specialized tasks over time.
Co-founder Ronak Malde told TechCrunch that the platform allows lawyers, doctors, and other specialists to improve models with little technical involvement, which he said could make deploying specialized AI faster and cheaper. He described the people using the platform as deciding "what good looks like for their organization" and turning that into model improvements. Harvey and Rogo are already using Trajectory to post-train open-source models to perform more like experienced lawyers or financial analysts, such as by reviewing specific contract types or valuing companies using particular methodologies. Other customers include Decagon, Mercor, and Nvidia. The company is now making its post-training platform publicly available.
Trajectory's approach shares similarities with recursive self-improvement, but co-founder Michael Elabd argues that keeping humans in the training loop is safer than autonomous self-improvement, which could produce unintended or dangerous outcomes. The company names Thinking Machines Lab, founded by former OpenAI CTO Mira Murati and behind the Tinker fine-tuning platform, as a direct competitor. Other players in automated post-training include Applied Compute, valued at $3.25 billion last month, and Prime Intellect.
Malde cited the team as the startup's largest advantage. He was among a handful of researchers Google DeepMind acqui-hired from Windsurf, Elabd also worked at DeepMind, and co-founder Arjun Karanam previously worked on generative foundation models at Apple. The company has also recruited researchers and product designers from OpenAI, Stripe, and Netflix. Karanam said the company has built an interface intended to make post-training accessible to non-technical customers.
The startup addresses a core limitation of current AI models, which do not learn from their mistakes without further training. Businesses adapting general models for specialized fields such as law and finance typically fine-tune them, a process that often requires internal research teams or costly consultants known as forward-deployed engineers. Trajectory aims to automate this work through continual learning. Companies connect the platform to their applications and start with an open-source model. The platform collects feedback on the model's performance and uses it to retrain the model, improving its handling of specialized tasks over time.
Co-founder Ronak Malde told TechCrunch that the platform allows lawyers, doctors, and other specialists to improve models with little technical involvement, which he said could make deploying specialized AI faster and cheaper. He described the people using the platform as deciding "what good looks like for their organization" and turning that into model improvements. Harvey and Rogo are already using Trajectory to post-train open-source models to perform more like experienced lawyers or financial analysts, such as by reviewing specific contract types or valuing companies using particular methodologies. Other customers include Decagon, Mercor, and Nvidia. The company is now making its post-training platform publicly available.
Trajectory's approach shares similarities with recursive self-improvement, but co-founder Michael Elabd argues that keeping humans in the training loop is safer than autonomous self-improvement, which could produce unintended or dangerous outcomes. The company names Thinking Machines Lab, founded by former OpenAI CTO Mira Murati and behind the Tinker fine-tuning platform, as a direct competitor. Other players in automated post-training include Applied Compute, valued at $3.25 billion last month, and Prime Intellect.
Malde cited the team as the startup's largest advantage. He was among a handful of researchers Google DeepMind acqui-hired from Windsurf, Elabd also worked at DeepMind, and co-founder Arjun Karanam previously worked on generative foundation models at Apple. The company has also recruited researchers and product designers from OpenAI, Stripe, and Netflix. Karanam said the company has built an interface intended to make post-training accessible to non-technical customers.
Go deeper than the headline
Research every firm in this story, track the ones that matter and get the daily briefing — free for 7 days.