Trajectory raised $45 million to let customers retrain their own models. Defining "good" Is the hard part. SeaPRwire - Trajectory has raised $45 million in a Series A round at a $300 million valuation, led by Sequoia, with Nvidia, Bessemer Venture Partners and Bain Capital Ventures participating. The company has 18 employees and a platform that until now has been used by a short list of design partners. It opened that platform to outside customers this week. The problem it sells against is specific. A law firm that wants a model to read its own contract templates has three options: hire researchers, rent consultants known in the industry as forward-deployed engineers, or accept a general model's guesses. Fine-tuning works, but somebody has to run it, and the people who know what a good answer looks like are usually the ones who cannot write a training script. The three options cost money in different ways. Hiring researchers is slow and permanent, renting consultants is fast and expensive, and accepting a general model's guesses stays cheap until somebody reads the output. Trajectory's product closes that loop. A customer starts from an open source model, connects the platform to the application where the model runs, and lets it collect signals on how each output performed. The platform then retrains the model on those signals. The customer keeps the weights and the loop, which is the part that separates this from an API subscription. None of that is a training problem. The hard part sits upstream of the training run: deciding which outputs were wrong, recording why, and doing it often enough that the model sees a pattern instead of noise. Firms that already review work in writing start with an advantage. Firms that review in conversation will find the platform asking them to formalize something nobody has ever written down. Ronak Malde, a co-founder, described the gap in terms a buyer can feel. A foundation model, he said, is like a very smart PhD student who starts a new job every week. "But what the economy values is someone who worked on the job for 30 years," he told TechCrunch. The analogy is doing work that the product description does not. A model that has read every contract in the world still has not read this firm's contracts, and no amount of scale substitutes for the corrections a reviewer makes on Tuesday afternoon. What those corrections are, and who supplies them, is where the difficulty moves. Feedback has to be labeled before it can train anything. A lawyer marking an answer wrong has to say why, in a form the platform can use, and a firm that cannot do that consistently will train a model to reproduce its own inconsistencies faster. Malde's framing puts that judgment on the customer by design. The people using Trajectory decide "what good looks like for their organization and can turn that into improvements for their model," he said. Early customers suggest where the demand sits. Harvey and Rogo, both legal and financial research tools, use the platform to post-train open source models on tasks such as reviewing a particular type of contract or valuing a company with a specific methodology. Decagon, Mercor and Nvidia are also customers. The list also shows what the platform is not. None of these companies sells a general assistant; each sells a tool that has to be right about a narrow task, which is the profile where a retrained open source model can hold its own against a larger general one on cost. Continual learning invites comparison with recursive self-improvement, the idea that a system improves its own capabilities without a person in the loop. Michael Elabd, another co-founder, argues the human keeps the process safer, since autonomous self-improvement can produce outcomes nobody intended. The distinction is not cosmetic: a model that retrains on its own outputs can drift, and the drift is hard to see from inside the loop that produced it. Competition is well funded. Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, ships the fine-tuning platform Tinker. Applied Compute was valued at $3.25 billion last month. Prime Intellect raised $130 million in July to help companies build and train their own agents. Post-training has become a category with capital behind it, and the pricing pressure that follows. Every one of those companies sells to the same buyer, and that buyer has not yet decided what the line item should be called. Trajectory's answer to that is its team. Malde was among the researchers Google DeepMind acqui-hired from Windsurf. Elabd also came from DeepMind. Arjun Karanam, the third co-founder, worked on generative foundation models at Apple. The company has since hired researchers and product designers from OpenAI, Stripe and Netflix. That last detail explains a choice that otherwise looks odd. A training company hiring product designers is unusual until you accept that the constraint is not the training run. It is the interface where a specialist who has never opened a notebook records what counted as a good answer, and does it for the hundredth time without giving up. Karanam makes the case directly: the interface is what makes post-training usable by people who are not researchers. Whether that holds up under a customer's own audit is the open question. A retrained model has to be versioned, rolled back and explained to whoever signs off on it, and a platform that makes retraining cheap also makes it easy to accumulate a stack of model versions nobody can account for. Malde names the team as the company's main advantage, which is also the answer a founder gives when the technology is not yet the moat. The test will come from customers who keep the loop running after the first model ships, and from whether the improvements survive a change in the underlying open source model. Proof, if it arrives, will look mundane: a customer running the same loop for a year, a model that measurably improves on the task it was tuned for, and a cost per finished task that falls. None of those numbers appear in a funding announcement, and none of them can be read off a valuation.

Trajectory opens self-serve platform for AI post-training feedback. The platform routes edits and corrections from AI products into recurring cycles. Its continual grading feature turns human preferences into repeatable scores. Published Nov 2, 2026, 8:13 PM EST Trajectory opened a self-serve platform Nov. 2 that routes edits and corrections from users of AI products into recurring post-training cycles, giving qualifying businesses a way to feed production feedback into model-refinement work. Discover more Bitcoin price prediction Bitcoin investment guide Ethereum blockchain development Businesses connect Trajectory's software development kit (SDK) to their products, where the platform captures signals such as changes users make to AI-generated output. Those signals feed post-training cycles that can run weekly or more often. Post-training adjusts a model after its initial development to shape its behavior for particular tasks. Its continual grading feature is designed to turn human preferences into repeatable scores. Its GS1 algorithm enables real-time learning from user actions. Trajectory's platform uses product feedback in post-training, while Together AI's Oct. 19 post-training service offers reserved GPU capacity and developer-selected training workflows. Both services focus on model customization, with different inputs and infrastructure for developers. The launch extends access beyond a small group of partners, allowing qualifying businesses to sign up directly. Trajectory's client data usage adheres to SOC 2 standards, a security and compliance framework, and user control is an operating principle. User corrections can be wrong, making feedback quality relevant to a system that uses those corrections as signals. Simon Yoon. Simon Yoon reports on blockchain technology for TokenPost. Send corrections or tips to [email protected].