Trajectory raises $45M to automate AI model post-training

Trajectory, an 18-person startup founded about a year ago by three Stanford classmates, has raised $45 million in a Series A round at a $300 million valuation. Sequoia led the round, with participation from Nvidia, Bessemer Venture Partners, and Bain Capital Ventures. The company announced the funding alongside the public availability of its post-training platform.

The company's pitch targets a structural limitation in current AI models. Whether developed by OpenAI, Anthropic, or built on open-source alternatives, today's models do not learn from their mistakes without further training. Businesses adapting these systems for specialized fields such as law and finance typically work around this by fine-tuning general models. That process usually requires hiring internal research teams or relying on expensive consultants known as forward-deployed engineers, or FDEs.

Trajectory wants to automate that work. The platform lets AI models learn from feedback and continuously improve at specialized tasks, without a standing team of machine-learning engineers.

What does the product actually do?

Trajectory's approach rests on a technique known as continual learning. A business starts with an open-source AI model and connects the Trajectory platform to its applications. The platform then collects feedback on how the model performs in real use. Trajectory uses that information to retrain the model, improving its performance on specialized tasks over time.

The intended users are not machine-learning researchers. They are the domain experts themselves. Co-founder Ronak Malde said lawyers, doctors, and other specialists can improve the models "near automatically" themselves. That capability, he argued, could make deploying specialized AI both faster and cheaper.

The customers already on the platform illustrate the use cases. Harvey and Rogo, two companies building AI for legal and financial work, use Trajectory to post-train open-source models so they perform more like experienced lawyers or financial analysts. The models learn from previous tasks, such as reviewing a particular type of legal contract or valuing a company using a specific methodology. Trajectory's other customers include Decagon, Mercor, and Nvidia — the chipmaker that also invested in the Series A.

The people using Trajectory decide "what good looks like for their organization and can turn that into improvements for their model," Malde said.

Why does the founder think today's models fall short?

Malde frames the problem with a comparison to hiring. Foundational models, he said, resemble a very smart PhD student who is always starting a new job. "But what the economy values is someone who worked on the job for 30 years," he told TechCrunch.

That gap between raw capability and accumulated job-specific experience is the market Trajectory is chasing. Fine-tuning has so far closed the gap through labor: internal research teams or FDEs who manually adjust models for a client's workflows. If continual learning removes that labor, the cost structure of deploying specialized AI changes. Nvidia's participation in the round also links the startup to the largest supplier of AI training infrastructure.

How does this relate to recursive self-improvement?

Trajectory's method shares similarities with recursive self-improvement, or RSI — the idea that AI systems can improve their own capabilities. That concept carries significant weight in AI safety debates, because an autonomously self-improving system could develop unintended or even dangerous outcomes.

Co-founder Michael Elabd draws a line between Trajectory's approach and fully autonomous self-improvement. He argues that keeping humans involved in the training process makes Trajectory's approach safer. Specialists define what counts as a good output, and the platform turns those judgments into model improvements. The system improves from human feedback rather than rewriting its own capabilities without oversight.

Who is Trajectory competing with?

Trajectory is not alone in trying to automate post-training. The company considers three startups direct competitors:

• Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, which built the fine-tuning platform Tinker.
• Applied Compute, valued at $3.25 billion last month.
• Prime Intellect, which helps companies build and train their own AI agents.

The presence of Murati's Thinking Machines Lab in the field signals how much attention post-training now receives. As base models from major labs commoditize, the tools that adapt those models to specific domains — and keep improving them after deployment — have become a distinct and contested market segment.

Why does the team matter so much?

Malde believes the startup's biggest advantage is its people. He was among the handful of researchers Google DeepMind acqui-hired from Windsurf. Elabd also worked at DeepMind. The third co-founder, Arjun Karanam, previously worked on generative foundation models at Apple.

Beyond the founding trio, Trajectory has recruited researchers and product designers from OpenAI, Stripe, and Netflix.

The inclusion of designers at an AI training company is deliberate. Karanam says the startup has developed an interface that makes post-training more accessible to its non-technical customers. The bet is that the bottleneck in specialized AI is not model quality but the usability of the tools that adapt models to expert workflows. If a lawyer can specify what a good contract review looks like, and the platform can translate that specification into a better model, the FDE becomes unnecessary.

What happens next?

Trajectory is now making its post-training platform publicly available, moving from private deployment with named customers such as Harvey, Rogo, Decagon, Mercor, and Nvidia to open access. The company enters that wider market with $45 million in fresh capital and a valuation that prices it at one-fifteenth the size of its competitor Applied Compute's $3.25 billion mark from last month. The test ahead: whether domain experts, given tools that promise near-automatic improvement, can produce models that rival those shaped by years of engineering labor.