Trajectory raises $45M so non-technical experts can post-train AI models themselves

Trajectory raises $45M Series A led by Sequoia

Trajectory, an 18-person startup founded about a year ago, has raised $45 million in a Series A round at a $300 million valuation, according to TechCrunch. Sequoia led the round, with Nvidia, Bessemer Venture Partners and Bain Capital Ventures also participating.

The company was started by three Stanford classmates: Ronak Malde, Michael Elabd and Arjun Karanam. Per TechCrunch, Malde was among the small group of researchers Google DeepMind acqui-hired from Windsurf, Elabd also worked at DeepMind, and Karanam built generative foundation models at Apple. The startup has since recruited researchers and product designers from OpenAI, Stripe and Netflix.

The bottleneck Trajectory is targeting

Today's leading models — whether from OpenAI, Anthropic or built on open source foundations — share a structural limitation: they cannot absorb lessons from their own errors without additional training. Companies adapting these models for specialized fields such as law and finance typically fine-tune them, but as TechCrunch points out, that usually means hiring internal research teams or paying for expensive consultants, often described as forward-deployed engineers.

Malde's framing captures the gap: a foundation model resembles a brilliant PhD graduate who begins a brand-new job every single day, whereas what organizations actually need is the equivalent of a colleague with decades of experience in that exact role.

Continual learning, with humans still in the loop

Trajectory's answer is an approach known as continual learning. A customer starts with an open source model and connects the platform to their applications, allowing it to gather feedback on how the model performs in real use. Trajectory then uses that information to retrain the model so it gets better at specialized tasks over time.

Malde told TechCrunch that lawyers, doctors and other specialists can improve the models near-automatically themselves, defining what counts as good performance in their organization and translating that judgment into concrete model improvements. Karanam said the company has built an interface specifically to make post-training accessible to non-technical users — which explains why product designers sit alongside AI researchers on the team.

The approach has surface similarities to recursive self-improvement (RSI), the notion that AI systems can enhance their own capabilities. Elabd argues that keeping humans involved in the training process makes Trajectory's version safer, since fully autonomous self-improvement could lead to unintended or even dangerous outcomes.

Customers and competitors

Harvey and Rogo already use the platform to post-train open source models so they perform closer to the level of seasoned lawyers or financial analysts, drawing on past work such as reviewing specific kinds of contracts or valuing companies with particular methodologies. Trajectory's other customers include Decagon, Mercor and Nvidia, and the company is now making its post-training platform publicly available.

The startup is not alone in trying to automate post-training. TechCrunch reports that Trajectory views Thinking Machines Lab — founded by former OpenAI CTO Mira Murati and the company behind the fine-tuning platform Tinker — as a direct competitor, alongside Applied Compute, which was valued at $3.25 billion last month, and Prime Intellect, which helps companies build and train their own AI agents.

Why it matters

Fine-tuning is one of the main bottlenecks standing between general-purpose models and AI that is genuinely useful in specialized professions. Today that work is gated behind ML engineers and costly consulting engagements; if domain experts can instead encode their expertise directly into models through feedback, deploying specialized AI becomes both faster and cheaper.

The round is also a signal about where investors see value accruing in the AI stack: not only in frontier models, but in the post-training layer built around them. Sequoia and Nvidia backing an 18-person company at a $300 million valuation — while well-funded rivals like Thinking Machines Lab and Applied Compute are already in the market — suggests the race to own that layer is well underway.