Trajectory raises $45M from Sequoia to let domain experts post-train open models

The round

Trajectory 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, TechCrunch reported. The company is 18 people and was founded about a year ago by three Stanford classmates with experience at Google DeepMind and Apple.

The same day, Trajectory made its post-training platform publicly available. It had been running with a small set of design partners in legal, financial and customer support work, the kind of jobs where the difference between a general model and a specialist one shows up in whether the output is usable at all.

The valuation is the interesting part of the headline. A $300 million price for a year-old, 18-person company with no disclosed revenue is a statement about where investors think the money in AI is going next: not into another frontier lab, but into the tooling that turns someone else's frontier model into a working product.

What continual learning means in practice

Whether developed by OpenAI or Anthropic or built on open source alternatives, current AI models share a limitation: they do not learn from their mistakes without further training. Businesses adapting them for specialised fields such as law and finance typically fine-tune general models, which usually requires either an internal research team or expensive consultants known as forward-deployed engineers.

Trajectory's answer is to automate that loop. A business starts with an open source model and connects the platform to its applications, which lets Trajectory collect feedback on how the model performs. That feedback is used to retrain the model so it improves at specialised tasks over time, and in the company's framing, lawyers, doctors and other specialists can drive the improvement themselves rather than filing a request with a research team.

Co-founder Ronak Malde described the problem with an analogy rather than a benchmark. Foundational models, he said, are like 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.

The customers named so far are a reasonable sample of that thesis. Harvey and Rogo are already using Trajectory's platform to post-train open source models so they perform more like experienced lawyers or financial analysts, learning from previous tasks such as reviewing a particular type of legal contract or valuing a company with a specific methodology. Trajectory's other customers include Decagon, Mercor and Nvidia.

The product, as the company describes it

Trajectory's own site pitches a single unified training platform: train, evaluate and deploy improvements without waiting for a lab or the next big model release. The claim that matters operationally is that it "works with any agent harness and takes just two code changes to get started", which is the sort of sentence that separates a developer tool from a consulting engagement.

Two of the named features describe what happens after launch. Continual Grading lets a customer define what good looks like on a subset of traces and apply that standard across all production data. Group Size 1 lets a model train on every production interaction instead of freezing at launch, which is the difference between a model that ships and a model that keeps moving.

Governance is packaged as a feature rather than a caveat. The company says it is SOC 2 certified, that every model update runs through the customer's eval suite and approval workflow before it touches production, and that each training run can be audited to see what the system learned and why it changed. Those controls are the commercial precondition for selling retraining into regulated work, where an unexplained model change is a compliance event.

The site also states cumulative funding of $60 million across a Series A and seed, consistent with the $45 million round reported by TechCrunch plus earlier seed capital. Trajectory's homepage quotes Nvidia chief executive Jensen Huang, speaking at GTC, describing Trajectory and Harvey post-training the Nemotron model into frontier-level capability for their use case.

The competitive set is not small

Trajectory is not the only company trying to automate post-training. It considers Thinking Machines Lab, founded by former OpenAI chief technology officer Mira Murati and the company behind the fine-tuning platform Tinker, a direct competitor, along with Applied Compute, which was valued at $3.25 billion last month, and Prime Intellect, which helps companies build and train their own AI agents.

That list is a reminder that the category has already attracted serious capital and that the technical moat is not obvious from the outside. If post-training becomes a commodity interface over open weights, the winner is whoever has the best evaluation loop and the least friction, not whoever trains the biggest model.

Malde's answer is the team. He was among the handful of researchers Google DeepMind acqui-hired from Windsurf, Michael Elabd also worked at DeepMind, and the third co-founder, Arjun Karanam, previously worked on generative foundation models at Apple. The startup has also recruited researchers and product designers from OpenAI, Stripe and Netflix, and Karanam argues that the interface is what makes post-training accessible to customers who are not researchers.

The safety framing, and where it gets awkward

Trajectory's approach shares some similarities with recursive self-improvement, the idea that AI systems can improve their own capabilities. Co-founder Michael Elabd argues that keeping humans involved in the training process makes Trajectory's approach safer, since autonomous self-improvement could lead to unintended or even dangerous outcomes.

That argument is worth taking seriously and worth testing. The capability Trajectory sells, a model that keeps learning from production interaction, is mechanically the same direction of travel that safety teams at the large labs describe when they talk about systems improving themselves. The difference is who decides what counts as an improvement: in Trajectory's product, a human customer defines the standard and approves the update. Our own coverage of Anthropic's S1A filing noted the company warning that a recursive self-improvement threshold may be near, which is precisely the regime this class of tooling operates at the edges of.

The commercial version of that debate will not be settled by positioning. It will be settled by whether customers in law, medicine and finance treat the approval workflow as a real gate or as a click.

What the money implies

A $300 million valuation for 18 people and a year of work implies a specific bet: that post-training spend grows faster than training spend, and that most companies end up with models tuned on their own data rather than renting a shared frontier model. That is the same bet behind the wave of inference infrastructure funding, except Trajectory sits one layer up, where the output is a better model rather than a cheaper token.

The risk is base-rate drift. If the big labs keep shipping models that already handle a law firm's contract review or a bank's valuation methodology, the value of owning a retrained open model falls. If they do not, and the gap between general and specialist stays wide, every serious operator eventually buys something like this.

There is also the question of what a valuation like this requires in return. Applied Compute reached $3.25 billion last month on a similar premise, which means Trajectory will be compared against companies with more capital and, in some cases, more compute. Being early and small is an advantage only while the product stays ahead of the interface the labs ship for free.

What to watch

The first signal is whether the named design partners publish measurable results, because a platform that makes models better is a claim that can be checked. The second is whether Trajectory keeps working when the base models underneath it change, since continual learning on an open source model is only as stable as the model's release cadence.

The third is pricing. Nobody has said what a retraining run costs a customer, and that number decides whether continual learning becomes routine infrastructure or another specialist tool that only well-funded teams can afford. Right now the company is selling a loop, and the loop is the interesting part of the AI stack that nobody has priced yet.