Trajectory made its post-training platform publicly available on November 2, 2026, offering organizations a way to connect AI applications, collect feedback and retrain open models for specialized work. The pitch centers on the people who understand the task: lawyers, financial analysts and other experts could steer model improvement without building a training system themselves, according to TechCrunch’s launch report.
The platform brings data collection, training, evaluation and deployment into one interface. Trajectory says customers control which data enters training and approve model updates before they reach production. That makes this a human-guided improvement workflow, with experts defining the target and reviewing what ships.
The launch accompanied a $45 million Series A at a $300 million valuation, led by Sequoia. The more consequential product question is what happens between a user correcting an AI answer and a new model reaching production. Trajectory’s published materials describe that process, while leaving important questions about performance, access and operating costs unanswered.
How Production Feedback Becomes Training Data
Most deployed models do not update their underlying parameters simply because someone corrects an answer. A correction might help within an ongoing conversation or be saved by an application, but changing the model itself requires a separate training process.
Trajectory aims to connect those stages. Its platform homepage describes capturing execution records, corrections, edits and re-prompts from an existing application, then using those signals to shape subsequent training.
The documented workflow has five parts:
- Connect the application. An SDK captures traces and user feedback from the existing agent or application.
- Choose useful data. The organization decides which interactions can be used for training.
- Define quality. Experts supply examples and grading criteria that describe successful work.
- Train and evaluate. The platform runs post-training and checks the resulting model against the organization’s evaluation suite.
- Approve deployment. A model update passes through the customer’s approval workflow before reaching production.
Trajectory says integration takes two code changes and works with any agent harness, meaning the software that coordinates an agent’s execution. Those are vendor claims about integration effort. They should not be read as a complete estimate for preparing data, designing evaluations or securing approval to use sensitive records.
The feedback distinction matters. An edit is evidence that someone wanted a different answer; it is not automatically a reliable statement of why the original answer was wrong. A retry could indicate a factual error, an awkward format or a user changing their mind.
For example, a lawyer might rewrite a contract summary because it missed an exception. Training on that correction could be useful. Treating every shorter rewrite as proof that shorter answers are better could teach the wrong lesson.
My view is that the strongest part of Trajectory’s proposition is the connection between expert judgment and training infrastructure. Capturing feedback is relatively easy to describe. Turning it into a dependable learning objective is the harder, more valuable job.
Experts Still Have to Define Success
Trajectory calls its approach owning a value function: deciding which outcomes the model should optimize for. Co-founder Ronak Malde told TechCrunch that customers determine “what good looks like for their organization.”
In practical terms, that could mean scoring a legal answer for identifying the relevant clause, explaining an exception and citing the supporting text. A financial workflow might require the model to follow a particular valuation methodology and show its assumptions.
The interface may reduce the engineering work needed to run training. It cannot decide those professional standards on the customer’s behalf. The responsibility shifts toward making expert judgment explicit enough to evaluate consistently.
The Public Platform and Its Research Previews
Trajectory’s homepage presents training, evaluation and deployment as the platform’s core workflow. It separately labels two features as research previews: Continual Grading and Group Size 1. That distinction limits how broadly readers should interpret the continual-learning promise.
| Capability | What Trajectory says it does | Published status |
|---|---|---|
| Core training workflow | Connect applications, collect feedback, train models, evaluate updates and deploy approved versions | Public platform launch reported by TechCrunch |
| Continual Grading | Apply standards defined on a subset of traces across production data | Research preview |
| Group Size 1 | Enable training from individual production interactions | Research preview |
Trajectory’s homepage directly confirms the preview labels. A launch analysis from alfw.ai also distinguishes the available base workflow from those preview capabilities.
Continual Grading addresses a potential bottleneck: experts cannot manually assess every production interaction. The proposed mechanism is to define quality on a smaller set of examples and extend that grading across more data. Its usefulness depends on how faithfully the automated grading preserves the expert’s standards.
Group Size 1 targets training from individual interactions. The name alone does not explain the training algorithm, update frequency or safeguards. It should not be interpreted as a promise that every user interaction immediately changes the production model.
There is also an access wrinkle. Despite the public-launch announcement, the opened homepage still displays a waitlist invitation. The evidence supports a public platform launch, but it does not confirm immediate, unrestricted onboarding for every organization. Prospective users should check access to the base product and each preview separately.
What the 8.3% Benchmark Claim Actually Means
Trajectory’s named users include Harvey and Rogo, which TechCrunch says use the platform to post-train models for legal and financial tasks. TechCrunch also lists Decagon, Mercor and Nvidia as customers.
Those relationships provide context for the product’s intended use. They are not, on their own, measurements of how reliably it improves models.
The most specific performance claim in the supplied reporting concerns Harvey’s LAB benchmark. According to alfw.ai’s account of the launch announcement, Trajectory says it worked with Nvidia and Harvey to improve a Nemotron Lightning model from a 0% to an 8.3% all-pass rate.
| Item | Company-reported claim |
|---|---|
| Evaluation | Harvey’s LAB benchmark |
| Starting all-pass rate | 0% |
| Post-training all-pass rate | 8.3% |
| Comparison model | Opus 5.5, reportedly at the same score |
| Cost comparison | 100 times lower cost |
| Speed comparison | 3.8 times faster |
These are company-reported results relayed through secondary coverage, not independently verified measurements. Harvey is both a customer and the source of the benchmark. The available account does not provide enough methodology to reproduce the cost and speed comparisons.
The pass-rate arithmetic is still useful. Moving from 0% to 8.3% is an 8.3-percentage-point improvement. A percentage increase from a zero baseline is undefined.
An all-pass measure is strict: an evaluated case must meet all its passing criteria. At the reported result, roughly 91.7% of cases did not achieve an all-pass outcome. That does not mean every answer in those cases was wholly wrong, and it does not reveal how close unsuccessful cases came to passing.
I would read the result as evidence for a narrow company claim: post-training improved this model on this evaluation. It cannot support a general conclusion that the model matches Opus 5.5 across legal work or is ready for unsupervised production use.
The economics need the same care. A claim of 100 times lower cost is useful only when buyers know what is counted: model serving, training, total workflow cost or some other scope. Likewise, “3.8 times faster” needs a defined timing measure. Neither figure can be carried straight into a budget for an unrelated deployment.
Frequently Asked Questions
3 questions
1What Does Trajectory’s Platform Do?
It connects applications to a workflow for collecting feedback, post-training models, evaluating changes and deploying approved updates.
