Trajectory, a San Francisco startup, launched a platform on Nov. 2, 2026 that lets domain experts post-train AI models on their own data, and announced a $45 million Series A led by Sequoia at a $300 million valuation. Post-training is the stage after a model's initial development, when it is tuned to perform better on specific tasks.
The company calls the product the first self-serve option for post-training on production data. That is Trajectory's own claim and has not been independently established, and access remains gated by a waitlist.
What the platform does
Trajectory argues that the people closest to a job, and not only machine-learning specialists, should shape how a model behaves. "The people closest to the work know what a better model looks like," the team wrote in its launch announcement.
Ronak Malde, the CEO and a co-founder, made a similar case to TechCrunch's Marina Temkin. Domain experts, he said, know "what good looks like for their organization and can turn that into improvements for their model."
The product is more specific than a generic no-code label suggests. Per Trajectory's documentation, users set the token budget, the maximum number of turns, the reasoning effort and the reinforcement-learning algorithm. Trajectory handles sandboxing and high-throughput rollouts, and it also supplies observability tooling.
Access, pricing and what remains in beta
The homepage offers a login and a waitlist form, not open signup. Documentation describes production learning, including the collection of production traces and telemetry, as a private beta. Two other features, Continual Grading and Group Size 1, are labeled research previews.
Pricing has not been disclosed. FounderMag, in its Nov. 2, 2026 launch coverage, noted the missing prices and the waitlist as well. The official pages reviewed show no free tier, no training-price schedule and no enterprise contract structure.
Nor do Trajectory's public pages specify its SOC 2 report type, the audit period, or the full terms of data retention. Those details tend to matter to the sort of regulated customers the company is courting.
Customers and early results
The announcement names Harvey, Rogo, Mercor and NVIDIA as organizations already training on the platform. Contract terms and revenue are undisclosed, and the launch materials give no figures for retention or the number of paying customers.
The company also reported that its Nemotron Lightning model rose from 0% to 8.3% all-pass on Harvey LAB, a result it said matches Opus 5.5 at 100 times lower cost and 3.8 times the speed. The announcement did not lay out how cost was calculated or how broad the task coverage was, and no independent replication turned up.
Gabe Pereyra, Harvey's president and co-founder, appears in a homepage testimonial: "With Trajectory, we're building toward agents that can carry the evolving domain expertise behind important legal work." The quote is undated.
Trajectory's own research comes with caveats. In a legal experiment on generated rubrics, the confidence intervals overlapped, and the authors said the result does not establish those rubrics as superior. A separate post on single-rollout learning states that it is not evidence that every production interaction immediately changes deployed weights.
Funding and team
Sequoia led the round, and NVIDIA, Bessemer Venture Partners and Bain Capital Ventures took part, according to the announcement. It did not say whether the $300 million valuation is pre- or post-money.
The company's website lists $60 million as "Series A + Seed," which squares with the $45 million round plus an earlier $15 million seed. WIRED reported that seed, led by Conviction at a $115 million post-money valuation. Bessemer, Radical VC and BoxGroup participated, as did the individual backers Jeff Dean and Fei-Fei Li.
Other databases tell a muddier story. Dealroom and Caplight recorded a $40 million Series A in August, with $55 million in total funding. Their sources do not explain the gap with the November figure, so the amounts should not be added together. Some outlets, Crypto Briefing among them, reported a total of $60 million to $65 million, which conflicts with the company's own $60 million.
TechCrunch reported that Trajectory has 18 employees. WIRED reported that the founders' previous employers include Windsurf and Google DeepMind, the latter including its robotics division. One founder also worked at Apple on Vision Pro.
How it compares with rivals
Thinking Machines' Tinker emphasizes programmable control over training and managed infrastructure. Trajectory, by each vendor's own description, stresses an integrated workflow for experts plus previews of production learning. Applied Compute also pitches helping companies own and improve their models. None of this amounts to a head-to-head comparison.
A bigger question was raised by Altis Research on Sept. 10, 2026. The firm questioned whether continual training on production data becomes a separately purchased platform or simply a feature bundled by inference providers.
What to watch
Trajectory has the backing and a roster of named customers. What it has not yet shown is revenue, pricing, or how quickly waitlisted teams get in.
Runloop CEO Jonathan Wall, in a June 10, 2026 release that predates the launch, put the company's wager this way: "Trajectory is betting that the next generation of AI products will be defined by their training loops, not just their model weights."
