Executive Summary
A company that wants lawyers and analysts to do the teaching has raised a large round. Trajectory announced a $45 million Series A led by Sequoia on November 2, at a valuation of $300 million. NVIDIA, Bessemer Venture Partners and Bain Capital Ventures came in alongside. Three Stanford classmates founded the company about a year ago, and it employs 18 people.
The product is a continual learning platform, and the report notes that the same announcement opened it up: "The startup is now making its post-training platform publicly available." A customer connects an open-source model to its own application, and the platform gathers feedback from real work and retrains the model on it. What the company puts forward is who leaves that feedback. The judgment comes from the non-technical specialist who does the work every day, rather than from a research team or an outside labeling vendor. Co-founder Ronak Malde compares a foundation model to "a very smart PhD student who's always starting a new job," and adds that "what the economy values is someone who worked on the job for 30 years."
Sections 1 through 4 follow what the TechCrunch report states and what else has been published. Sections 5 and 6 ask who inside a company writes that standard and whose property the record becomes once a lawyer's daily judgment turns into training data. That part is this article's reading.
Key Numbers
Four numbers carry the announcement. The first three size the company, and the last belongs to a rival in the same market that was priced first.
Source: TechCrunch (2026-11-02)
$45 million
Series A raised
Led by Sequoia, with NVIDIA, Bessemer and Bain Capital Ventures taking part
$300 million
Valuation in this round
Set about a year after founding, at 18 employees
5
Customers named publicly
Harvey, Rogo, Decagon, Mercor and NVIDIA
$3.25 billion
Applied Compute's valuation
Set in October 2026, more than ten times Trajectory's
One Name Sits on Both the Investor List and the Customer List
The announcement came on November 2. Sequoia led the round, with NVIDIA, Bessemer Venture Partners and Bain Capital Ventures taking part. The round was $45 million at a $300 million valuation. The founders are three Stanford classmates. Ronak Malde was among the handful of researchers Google DeepMind acqui-hired from Windsurf, and Michael Elabd also worked at DeepMind. Arjun Karanam built generative foundation models at Apple. A little over a year old, the company has 18 employees and has recruited researchers and product designers from OpenAI, Stripe and Netflix.
What catches the eye is the overlap between two lists. NVIDIA put money into this round and is also one of the five customers the report names. The other four are Harvey, a legal AI company; Rogo, a financial AI company; Decagon, which builds customer-support agents; and Mercor, a marketplace that connects domain experts to AI labs. That last name comes back later.
What the company sells is a continual learning platform. A customer connects an open-source AI model to its own application, and the platform collects feedback on how the model performs during real use, then retrains the model on that feedback. Each turn of the loop makes the model fit that company's particular work a little better. Until now the platform ran inside the customers named above, and on the day of the funding news the company said it is making the post-training platform publicly available. The diagram below is one turn of that loop.
The Model Starts Its First Day Over Every Time
The analogy Malde used in the report shows the company's premise in one stroke. A foundation model is like "a very smart PhD student who's always starting a new job." It tests well and has no idea how this particular organization works. "But what the economy values is someone who worked on the job for 30 years," Malde added.
Strip the analogy away and what remains is the absence of accumulation. The report sets the same premise in its opening line: whether a model comes from OpenAI or Anthropic or is built on an open-source alternative, current models do not learn from their mistakes without further training. However good the model you attach, what happened at the company yesterday does not survive inside the model today. Point out the same thing ten times and the eleventh answer repeats the mistake. That is the gap continual learning claims to fill.
Two cases in the report spell out what that accumulation consists of. At Harvey, lawyers tune the model to do better at reviewing a specific type of legal contract. At Rogo, analysts teach it to value a company according to a set methodology. Neither is a push for general capability. Both move work products and working habits a firm already owns to the model side.
One distinction is worth keeping here. Picking a good model and fitting that model to your own work are two different problems. The first belongs to the model companies. The second belongs to the organization using the model, and inside that organization the person who knows what a job well done looks like is the practitioner, not the researcher.
The Quality Bar Passes from the Research Team to the Desk
Describing the product, Malde said domain experts such as lawyers or doctors can define "what good looks like for their organization and can turn that into improvements for their model." The report then writes in its own voice that with Trajectory, lawyers, doctors and other specialists can improve the models near automatically themselves. The whole weight of that sentence sits in its subject, and the subject has moved to the specialist.
Other people have filled that seat in post-training so far. Whether the data was preference pairs for reinforcement learning or worked examples for supervised training, the verdict on which answer is better came from trained labelers following guidelines a research team designed. A lawyer who took part was usually hired as a labeler, not sitting at their own desk working their own cases.
In the structure Trajectory proposes, that order flips. The lawyer stays in the chair where the contract review happens, and the act of editing, rejecting or approving the model's draft is itself the feedback. The line between a work record and training data disappears. That is convenient and unsettling at the same time. No dedicated team has to define the quality standard, which also means the model changes without anyone writing that standard down.
In one sentence: the authority to shape a model's character moves from the research organization to the business unit. Responsibility travels with that authority, and the report does not say who carries it or how.
Buying Expert Hours or Using the Ones You Already Pay For
The diagnosis that domain expertise decides model quality is not new in itself. A company that started from the same diagnosis and went the opposite way has already grown large. Mercor assembles lawyers, doctors, former investment bankers and similar specialists, connects them to frontier AI labs, and runs post-training work on their behalf. As reported in October 2025, the pool held more than 30,000 experts at an average rate of about $85 an hour, and the company was valued at $10 billion at the time. That is the path of buying expert time to manufacture training data.
Trajectory stands at the other end. Instead of hiring new experts, it takes the work an already-employed expert finished today and uses it as signal. This is also why the two do not look like straightforward rivals. Mercor appears on the customer list Trajectory published. A company selling expert data became a customer of a company that pipes expert feedback into models.
The other names aimed at this market stand on the engineering side. The report lines up Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, and its fine-tuning platform Tinker; Prime Intellect, which helps companies build and train their own AI agents; and Applied Compute, valued at $3.25 billion in October 2026. Trajectory's valuation is less than a tenth of Applied Compute's. The table below puts the four approaches in one place.
| Who | Who writes the quality standard | What they sell |
|---|---|---|
| Trajectory | Non-technical specialists the customer already employs | A platform that collects work feedback and runs the retraining |
| Mercor | Outside experts brought in on an hourly rate | Expert labor for post-training and the data they produce |
| Tinker (Thinking Machines Lab) | The customer's researchers and engineers | A fine-tuning platform |
| Prime Intellect | The customer's researchers and engineers | An environment for building and training your own agents |
The middle column is where Trajectory places its bet. The contest to build better tools for engineers already holds several well-funded players. Sidestepping that contest, the company hands the controls to a group that has never touched the training process.
Who Writes the Standard and Who Owns the Record
From here on, none of this is in the report. These are the things an organization should settle for itself before it signs anything.
5.1The Ten Minutes Nobody Counts
Leaving a record that says this answer is good and that one is not comes at a price. The ten minutes a lawyer spends fixing a model draft are ten minutes not spent on another matter. If those ten minutes pile up as a company asset, they are work rather than a favor. Yet this labor shows up in no performance metric. Nobody's review form carries a line that reads "taught the model," and on a busy day that is the first line to go. Before the question of whether the tool works comes the question of whose job description this gets written into.
5.2Where the Feedback Belongs When the Contract Ends
The report does not touch data ownership. That silence is itself a list of questions for the contract stage. Does the verdict a customer's lawyer leaves go only into that customer's model, or does it become an asset the platform can apply for other customers? When the contract ends, can the customer walk away with the accumulated feedback and the weights it trained? Starting from an open-source model makes these questions easier but does not answer all of them. You may get the weights and still find the move impractical if the pipeline that produced them and the evaluation criteria stay on the platform's side.
5.3What Keeping a Human in the Loop Asks of That Human
The safety question was raised by the report, not by the company. It notes that Trajectory's approach to continual learning "shares some similarities with recursive self-improvement (RSI), the idea that AI systems can improve their own capabilities," and then gives co-founder Elabd's answer. He argues that keeping humans involved in the training process makes this safer than autonomous self-improvement, which can produce unintended or dangerous results. The distinction holds. Accepting it, though, also means watching where the weight of safety moves. The human inside that loop is a lawyer with fifty contracts to clear, not a researcher. For the safeguard to work, that lawyer has to leave careful verdicts even on a busy day. A design that keeps a human in the loop is also a design that adds one more task to a human.
Why Pebblous Is Watching This Round
The $45 million is not the heart of this news. What Pebblous looks at is that feedback-as-data has hardened into product structure. Until now, AI-Ready Data has mostly meant tidying material before it goes into training. On a continual learning platform, the verdicts and corrections produced while the model runs become the training material for the next version. The object of data quality management widens from input to operation.
Seen that way, a familiar problem comes back. It is old news that two labelers judging the same case differently will make a model wobble. In a structure where twenty practitioners each send signals by their own standard, no guideline may exist anywhere to steady that wobble. If lawyer A prefers conservative wording and lawyer B prefers concise wording, what the combined model learns is a result nobody designed.
An organization weighing a platform like this would do well to write down three things first.
- Does a written standard for a good answer exist in our work? Without one, the signal going to the model is an average of standards that differ by person
- Whose job is it to write and maintain that standard? Work that appears on no review form disappears in a busy quarter
- At the end of the contract, can we take back the accumulated feedback and the weights trained on it? If so, in what format do we receive them
How practitioner judgment becomes data is a thread this blog has followed before. One example is our piece on an Anthropic experiment that measured which way an expert's judgment tilts when an AI works alongside them. What that experiment showed is that experts do not know which way they were pushed. In a structure that takes those verdicts straight as a training signal, writing the standard down first matters more.
Thank you for reading this far. The $45 million and $300 million figures, the 18 employees, the five customers, the founders' backgrounds and the competitor names come from the TechCrunch report of November 2. Mercor's expert pool, hourly rate and $10 billion valuation come from October 2025 coverage and should be read as of that date. Sections 5 and 6 are this article's reading, carrying those facts over to the work of an organization. If you are weighing a structure that feeds practitioner judgment into a model, we would like to hear who is writing the standard.
Pebblous Data Communication Team
November 4, 2026
References
Primary Report
- 1.Temkin, M. (2026). "Trajectory raises $45M to make AI model training accessible to nontechnical experts." TechCrunch, 2026-11-02.
Industry Background
- 2.Thinking Machines Lab. "One Year of Tinker."
- 3.Prime Intellect. "Introducing Lab."
- 4.Wikipedia contributors. "Mercor." (The expert-pool size and valuation reflect October 2025 coverage; the article marks that date in the text.)