Trajectory Opens Platform for Companies to Post-Train AI Models on Their Own Production Data
The Sequoia-backed startup says user corrections and edits can feed weekly model updates, with every change gated by the customer's own evals.
Published
Last updated / Fact-checked on
Trajectory, a San Francisco continual-learning startup, opened a platform on November 2 that lets companies post-train AI models on their own production data, Crypto Briefing reported. The company is making the platform publicly available, TechCrunch reported, alongside a $45 million Series A. The idea: every time a user corrects an AI's output, that fix becomes training material.
Photo: RonaldCandonga / Pixabay
Trajectory at a glance
| Item | Detail |
|---|---|
| Founders | Ronak Malde (CEO), Arjun Karanam, Michael Elabd |
| Launch | November 2, 2026 |
| Integration | SDK; "two code changes," per Trajectory |
| Update cadence | Weekly or more often, per Crypto Briefing |
| Funding | $15M seed; $45M Sequoia-led Series A at $300M valuation |
| Named customers | Harvey, Rogo, Decagon, Mercor, Nvidia, Clay |
Sources: TechCrunch, Trajectory, Crypto Briefing, Dealroom.
How the platform works
According to Trajectory's website, the platform works with any agent harness and takes "just two code changes to get started." A lightweight SDK captures traces, corrections, edits and re-prompts from production.
Teams can then bring examples of good work, run training, compare their model against leading frontier models and push updates live with a click, the company says.
Trajectory stresses control. "Every model update runs through your eval suite and approval workflow before it touches production," its website says. Customers decide which data trains their models and can audit every training run.
Crypto Briefing said post-training cycles can run weekly or more often, and that Trajectory says client data handling follows SOC 2 standards.
The research behind it
Trajectory published two research notes on November 2. Both are listed as research previews on its platform.
Continual Grading tackles tasks with no clear right answer, such as legal analysis or customer support. It learns a "metagrader" from a small set of expert-graded examples, then writes grading criteria for new tasks. On the τ³-Bench customer-service benchmark, a metagrader built from 60 labeled tasks raised a model's score from 80.0% to 84.9%, versus 87.8% using the benchmark's own graders, Trajectory reported. The authors note the search can overfit to its examples.
You Only Rollout Once introduces GS1, short for "Group Size 1." It is designed to learn from a single attempt at each task, without replaying tasks or training a separate critic model. That matters because live customer interactions cannot be rerun. On the LABBench2 benchmark, GS1 raised held-out pass rate from 63% to 68%, the company said.
These are company-published results, not peer-reviewed papers.
Funding and customers
Trajectory raised a $15 million seed round led by Conviction in June at a $115 million post-money valuation. Its Series A totals $45 million at a $300 million valuation, led by Sequoia with Nvidia, Bessemer Venture Partners and Bain Capital Ventures, TechCrunch reported. In August, The Information had reported a $40 million Sequoia-led round at that valuation, according to Dealroom and Crypto Briefing.
Together, the two rounds total $60 million. The startup has 18 employees, according to TechCrunch. Angel backers listed on Trajectory's site include Jeff Dean and Fei-Fei Li.
Dealroom said Malde and Elabd previously worked as Google DeepMind researchers and Karanam as an Apple researcher. TechCrunch said Harvey and Rogo use the platform to post-train open-source models, and named Decagon, Mercor and Nvidia as other customers. Trajectory's site also lists Clay. Its site quotes Nvidia CEO Jensen Huang at GTC: "Here's Trajectory and Harvey using Nemotron, post-trained it and created their own frontier-level capability."
Why it matters
For companies running AI agents, Trajectory is pitching a way to make models improve from everyday use rather than staying fixed after deployment. The eval-and-approval gate addresses the obvious risk of shipping untested changes.
Crypto Briefing flagged two open questions: training on production data raises privacy and consent stakes, and user corrections are not always right.
What to watch
- Public uptake. Who adopts the newly public platform beyond its early enterprise customers.
- Independent results. Whether customers report measurable gains outside Trajectory's own benchmarks.
- Data governance. How enterprise buyers judge SOC 2 claims and controls on training data.
More coverage is on our Artificial Intelligence hub.
How we reported this
We checked product claims against Trajectory's website and its two November 2 research notes, and funding against TechCrunch, with earlier reports from Dealroom and Crypto Briefing citing The Information. Benchmark figures are Trajectory's own and are not peer reviewed. We did not test the platform.
Published and fact-checked November 2, 2026.
Sources
- Trajectory, website: https://www.trajectory.ai/
- Trajectory, "Continual Grading" (Neil Kale), Nov. 2, 2026: https://www.trajectory.ai/field-notes/continual-grading
- Trajectory, "You Only Rollout Once" (Nikhil Vemuri and Dian Ang Yap), Nov. 2, 2026: https://www.trajectory.ai/field-notes/you-only-rollout-once
- Crypto Briefing, Diego Almada Lopez, Nov. 2, 2026: https://cryptobriefing.com/trajectory-launches-self-serve-ai-post-training-platform/
- Crypto Briefing, funding report: https://cryptobriefing.com/trajectory-sequoia-funding-300m-valuation/
- TechCrunch, Marina Temkin, "Trajectory raises $45M to make AI model training accessible to nontechnical experts," Nov. 2, 2026: https://techcrunch.com/2026/11/02/trajectory-raises-45m-to-make-ai-model-training-accessible-to-nontechnical-experts/
- Dealroom, "Trajectory raises $40M Series A at $300M valuation," Aug. 11, 2026: https://dealroom.co/news/144435-trajectory-raises-40m-series-a-at-300m-valuation/
DailyMech used AI tools to help research and draft this article; it was fact-checked by our editors.