What did Harvey and Rogo just do?

On 2 November 2026, TechCrunch reported that Harvey, in legal, and Rogo, in finance, post-train open models on Trajectory's platform so that the models perform more like experienced lawyers and financial analysts. The models learn from past tasks, such as reviewing a particular type of contract or valuing a company with a specific method.

Trajectory's product, as TechCrunch describes it, starts from an open model, connects it to a company's applications, collects feedback on its outputs and retrains it over time. Lawyers, doctors and other specialists can improve the model themselves with little engineering help, co-founder Ronak Malde said. Trajectory has a $45 million Series A, according to TechCrunch's 2 November 2026 report. Malde told TechCrunch that "what the economy values is someone who worked on the job for 30 years".

TechCrunch does not say where those past tasks come from. Both are AI-first challengers in categories incumbents have served for decades, and for a product company the nearest supply of past tasks is the work its own users do in the product.

Who will train a model for you now?

By the first days of November 2026, training a model on your own work was a purchase from several vendors. Trajectory's platform is publicly available, TechCrunch reported, and it named three competitors: Thinking Machines Lab with its Tinker service, Applied Compute, and Prime Intellect. On 3 November 2026 Snowflake tied Cortex Training, in public preview, to the gateway that captures agent traces and logs, and its blog calls the result a continuous feedback loop for model customisation. Cortex Training itself was first announced in June 2026. It offers supervised fine-tuning, reinforcement learning, distillation and continued pre-training on open-weight families such as Qwen, DeepSeek, Mistral and Gemma.

• $45M

  Series A for Trajectory, the post-training platform Harvey and Rogo use, according to TechCrunch.

  Source: TechCrunch, 2 Nov 2026
• 18

  People at Trajectory when TechCrunch reported the round. A team this size now sells post-training to Harvey and Rogo.

  Source: TechCrunch, 2 Nov 2026
• ~$300K

  What Prime Intellect's Will Brown argued an internal model should cost a company, as reported.

  Source: startuphub.ai, 2 Nov 2026

Snowflake's press release says its gateway observability records agent activity in the customer's own Snowflake account, and its product page promises to "capture agent traces, logs and model interactions". So if a customer routes the agents in your product through its own gateway, the record of work done in your product could land next to the customer's training service. Snowflake's announcements do not mention third-party SaaS products. Where each of your features writes its trace is still a product decision, and today it is mostly made by default.

What is the one input none of them sells?

Every one of these training services, from Trajectory's platform to Snowflake's Cortex Training, runs on the same input: feedback on outputs a model produced during real work. Trajectory collects it from the applications its customers connect, and Malde describes the expert's job as defining what good looks like for their own organisation. Snowflake's loop starts from the traces its gateway captures. In each case the vendor brings GPUs and a training recipe and waits for the customer to bring the feedback.

For a vertical SaaS company, that work happens in your product. The claim an adjuster overrode, the clause a reviewer rewrote and the reconciliation a controller fixed all happen on your screens, against your data model, by people who have done the job for years. Will Brown of Prime Intellect made a related point in a talk reported by startuphub.ai on 2 November 2026: rules belong in code, while preferences and heuristics belong in data, where they compound once the same signal has repeated many times. Corrections to the same kind of claim, made hundreds of times a month by the same team, are that kind of signal.

The claim an adjuster overrode and the clause a reviewer rewrote happen on your screens, against your data model.

What happens to a correction in your product today?

Take a claims workflow in an insurance product. A new claim arrives with a policy, a loss report and a set of photos. The AI feature reads them and proposes a reserve and a coverage decision. The adjuster has seen this pattern before: the photos understate the damage, and the policy carries an endorsement the feature did not weigh. She raises the reserve, keeps the coverage decision, types two lines explaining why, and saves.

In most products the saved record now holds the final reserve and little else. The proposal she overrode is gone, or sits in an application log that nobody joins back to the claim. The documents behind the proposal are attached to the claim file, with nothing recording which version the feature read at that moment. Her note lives in a free-text field. What happened next, whether the claim paid at that reserve, reopened or went to dispute, arrives months later in another table with nothing tying it to her correction.

A training example needs four parts, kept together and linked to the same run: the input and context the proposal used, what the system proposed, what the expert changed and why, and what happened next. Once they are stored apart, or overwritten at save, they cannot be put back together later. A product built around a model can record all four from its first release, because its own model learns from them.

One correction, two records

The same adjuster correction in a claims workflow, as most products store it and as a training set needs it

Why does a quarter's delay matter?

A model can be trained in any month. A correction that was not recorded in March cannot be recovered in September, because the proposal it corrected and the context behind it no longer exist in a form any trainer can use. Each month the challenger's product records and yours overwrites is a month its model gained examples and yours gained none.

Choosing a trainer is a purchase you can make next year, and the evidence that domain training pays off is already public in the published results from domain-trained open-weight models. Two years of corrections take another two years to earn.

What should you check this quarter?

The check takes a week and needs no vendor.

Where SynOS fits

SynOS runs inside your product, white-labelled, inside your own system boundary. Every agent run is recorded with the context it used from your own systems, and when an expert corrects an output, the correction is kept with that run. The record accumulates on your own infrastructure in the shape a training set is built from, for a model you own, whichever trainer you choose later. Agents built on it rehearse against your real systems first, with writes recorded rather than executed and every run scored, so the record starts before anything ships to a customer.

Questions