Trajectory Closes $45 Million Round to Automate Specialised AI Training

The Expensive Gap Between General Models and Specialised Work

Large language models arrive with impressive general capabilities but no memory of their failures. A legal AI that misreads a clause in a merger agreement will make the same mistake tomorrow unless someone retrains it. Fixing that gap has become a lucrative niche for forward-deployed engineers, consultants who embed inside enterprises to fine-tune models for tasks like contract review or equity valuation. The hourly rates are high, the timelines stretch for months, and the work remains inaccessible to the lawyers, doctors and analysts who actually understand the domain.

Trajectory, an 18-person startup founded roughly a year ago by three Stanford alumni with backgrounds at Google DeepMind and Apple, has raised $45 million in Series A funding at a $300 million valuation to automate that post-training process. Sequoia led the round, with participation from Nvidia, Bessemer Venture Partners and Bain Capital Ventures.

The company's pitch is straightforward: connect an open-source model to your application, let Trajectory collect performance feedback, then watch the platform retrain the model to improve at your specific tasks. No PhD required, no consultant retained.

Continual Learning for Vertical Applications

Co-founder Ronak Malde frames the problem as a mismatch between what foundational models offer and what enterprises need. A base model resembles a brilliant graduate student rotating through departments, he argues, whilst businesses value the judgement that comes from three decades on the job.

Trajectory's platform addresses this through continual learning. Once integrated, it monitors how the model performs in production, whether that means drafting securities filings, triaging patient symptoms or answering customer support queries. The system then uses that feedback to retrain the model, sharpening its performance on the tasks the organisation cares about.

Legal AI provider Harvey and financial intelligence platform Rogo are among the early adopters. Both companies use Trajectory to post-train open-source models so they behave less like generalists and more like associates who have reviewed hundreds of credit agreements or built dozens of discounted cash flow models.

Other customers include Decagon, Mercor and Nvidia itself. The startup made its platform publicly available this month.

The Safety Debate Around Self-Improving Systems

Trajectory's approach shares conceptual ground with recursive self-improvement, the notion that AI systems might eventually enhance their own architecture without human oversight. Co-founder Michael Elabd, who also spent time at DeepMind, insists the startup's design keeps humans in the loop. Organisations using the platform define success criteria and approve training runs, rather than handing the model full autonomy to rewrite itself.

That distinction matters in a regulatory environment where legislators and standards bodies are scrutinising autonomous capability gains. Elabd believes involving domain experts at each step reduces the risk of models drifting toward unintended behaviours, a concern that has shadowed recursive self-improvement research since the concept emerged.

A Crowded Post-Training Market

Trajectory is far from alone in chasing the post-training opportunity. Thinking Machines Lab, founded by former OpenAI chief technology officer Mira Murati and the team behind fine-tuning platform Tinker, targets a similar customer base. Applied Compute, valued at $3.25 billion last month, offers infrastructure for companies building custom agents. Prime Intellect provides tooling for organisations that want to train models from scratch rather than adapt existing ones.

Malde acknowledges the competitive pressure but points to the team's pedigree. He was among a small group of researchers Google DeepMind acqui-hired from Windsurf. Elabd also worked at DeepMind, whilst third co-founder Arjun Karanam previously built generative foundation models at Apple. The startup has since recruited talent from OpenAI, Stripe and Netflix, spanning both research and product design.

That emphasis on design might seem incongruous for a company focused on training infrastructure. Karanam argues it is essential. The platform's interface is meant to let non-technical users, lawyers and financial analysts among them, define what good performance looks like and translate that into model improvements without writing code or understanding backpropagation.

What Trajectory Tells Us About Enterprise AI Economics

At Opentechwire, we have tracked dozens of vertical AI startups over the past 18 months, and a pattern has emerged: the companies that survive the trough of disillusionment are those that solve workflow problems, not those that promise magical reasoning. Trajectory's $300 million valuation suggests investors believe continual learning is a workflow problem worth solving at scale.

The economics are compelling. If a mid-sized law firm can train a model to handle routine due diligence without hiring a team of forward-deployed engineers, the unit cost of legal AI drops sharply. That could accelerate adoption in sectors like healthcare and financial services, where regulatory constraints and liability concerns have kept deployment cautious.

But the sustainability of Trajectory's model depends on whether open-source foundational models remain competitive with closed alternatives from OpenAI, Anthropic and Google. If proprietary models pull far enough ahead in baseline capabilities, enterprises may decide the customisation overhead is not worth it and simply pay for API access to the frontier. Trajectory's bet is that open-source models, enhanced through domain-specific post-training, will offer better performance per dollar for specialised tasks than general-purpose APIs ever could.

The next 12 months will test that thesis. If Harvey, Rogo and others using Trajectory demonstrate measurably better outcomes than competitors relying on off-the-shelf APIs, the continual learning narrative gains credibility. If the performance gap narrows or the operational burden of managing training pipelines proves too high, the $45 million round may look optimistic in hindsight.

For now, the startup has capital, customers and a clear value proposition. Whether it can turn that into durable competitive advantage in a market where Applied Compute, Thinking Machines Lab and Prime Intellect are all raising at aggressive valuations remains the open question.