Trajectory raises $45M to bring model post-training to non-technical experts

Trajectory announced on November 2 that it raised a $45 million Series A at a $300 million valuation, led by Sequoia with participation from Nvidia, Bessemer Venture Partners and Bain Capital Ventures. The roughly one-year-old startup is also making its post-training platform publicly available.

The problem it targets is concrete. Whether a model comes from OpenAI, Anthropic or the open-source ecosystem, it does not learn from its mistakes without further training. Businesses adapting models to fields such as law and finance usually fine-tune a general model, which often means hiring an internal research team or paying for expensive forward-deployed engineers.

Trajectory's answer is continual learning. A customer starts with an open-source model, connects the platform to its applications so it can collect feedback on how the model performs, and Trajectory retrains the model on that information so it improves at specialized tasks over time. Co-founder Ronak Malde compares foundation models to a very smart PhD student who is always starting a new job, while what the economy values is someone who has worked the job for 30 years.

Customers are already using it that way. Harvey and Rogo post-train open-source models through Trajectory so they behave more like experienced lawyers or financial analysts on tasks such as reviewing a particular kind of legal contract or valuing a company with a specific methodology. Decagon, Mercor and Nvidia are also customers.

The team is the pitch. The three co-founders are Stanford classmates with stints at Google DeepMind and Apple: Malde was among the researchers DeepMind acqui-hired from Windsurf, Michael Elabd also worked at DeepMind, and Arjun Karanam previously worked on generative foundation models at Apple. The 18-person company has also recruited researchers and product designers from OpenAI, Stripe and Netflix.

Trajectory draws a deliberate line between its approach and recursive self-improvement. Elabd argues that keeping humans in the training loop makes the system safer, since autonomous self-improvement could produce unintended or even dangerous outcomes. In practice, the people using Trajectory decide what good looks like for their organization, and the platform turns that judgment into model improvements.

Competition is crowding in. TechCrunch notes that Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, and its fine-tuning platform Tinker are direct competitors, alongside Applied Compute, valued at $3.25 billion last month, and Prime Intellect, which helps companies build and train their own agents.

The round shows post-training turning from a capability owned by a few research teams into a purchasable layer. The real test comes after opening the platform up: whether teams without strong labeling and evaluation practices can reliably convert business feedback into better models. If they can, the cost and lead time of deploying specialized models get repriced.