AI model training has mostly been a job for research labs with big GPU budgets. A startup called Trajectory wants to change that. On November 2, 2026, it announced a $45 million Series A. Its goal is to let lawyers, analysts and other experts improve open source models using feedback from daily work.

I'm Nitin, and I lead project coordination at Debugged Pro. Clients keep asking us one question: should we train our own model? This funding round is a good moment to answer it honestly. Here is what Trajectory actually announced, and what it means for agencies and businesses in India and the USA.

Trajectory co-founders, the startup that raised $45M for AI model training — Trajectory's three co-founders. Image: Trajectory, via TechCrunch

What Trajectory Announced About AI Model Training

The news broke in TechCrunch's report on Trajectory's $45M raise. I also read the company's own launch post to check the details. Here are the verified facts:

The money. A $45 million Series A at a $300 million valuation, led by Sequoia. Nvidia, Bessemer Venture Partners and Bain Capital Ventures also took part.
The team. An 18-person company founded about a year ago by three Stanford classmates.
The founders. Ronak Malde and Michael Elabd both worked at Google DeepMind. Arjun Karanam worked on generative foundation models at Apple.
The customers. Harvey and Rogo use it to post-train open source models. Decagon, Mercor and Nvidia are also customers.
The launch. The post-training platform is now publicly available as a self-serve product.

The problem it is trying to solve

Today's models, open or closed, don't learn from their mistakes unless someone trains them again. Businesses in fields like law and finance usually fix this by fine-tuning a general model. That often means hiring a research team or paying expensive consultants, known as forward-deployed engineers.

Malde put it simply. He compared a foundation model to a very smart PhD student who keeps starting a new job. "But what the economy values is someone who worked on the job for 30 years," he told TechCrunch.

How Continual Learning Works in Plain Words

Trajectory calls its approach continual learning. A business starts with an open source model and connects the platform to its app. The platform collects feedback on how the model performs, then uses that feedback to retrain it.

Think of a support bot. A human agent fixes its answer, or escalates a case the bot got wrong. That correction becomes a lesson. Over time, the model picks up the team's own standards instead of staying frozen at launch.

According to Trajectory's launch post, the platform covers these steps:

Train and deploy. One click to train and one to deploy, without managing sandboxes or infrastructure.
Continual Grading. A research preview that learns from a few examples of your judgment and applies it across production data.
Group Size 1. Another research preview that trains on real production interactions so the model keeps improving after release.

The company says getting started takes about five lines of code. It also says it works with any agent harness.

What the company claims, and how to read it

Trajectory shares some early results. It says one customer trained more than eight models for complex agent tasks in two weeks. It also says a small Nvidia Nemotron model matched Opus 5.5 on Harvey's legal benchmark at 100x lower cost.

These are company claims, not independent tests. I find them useful as a direction, not as a promise. Your results will depend on your task, your data and how good your feedback is.

Why AI Model Training Matters for Agencies

For years, the agency playbook has been simple. Pick a hosted model, write good prompts, add retrieval, and ship. That still works for most projects. But a few pressures are pushing teams to look at training:

Token bills keep climbing. Trajectory's own post points to rising costs for every company building on AI.
Generic models miss context. They don't know a client's tone, rules or edge cases unless you keep reminding them.
Clients want ownership. Some prefer a model they control over one they rent from a single lab.
Open models are getting better. A smaller trained model can sometimes do a narrow job well at a lower price.

We wrote about the cost side in our post on ChatGPT usage limits and shared token pools. When usage caps tighten, a cheaper model for repeated tasks starts to look attractive.

Paper plane illustration from Trajectory's AI model training platform launch post — Artwork from Trajectory's platform launch post. Image: Trajectory

What it means for teams in India and the USA

Indian agencies already do a lot of AI integration work for clients abroad. Tools like this could open a new service line: setting up feedback loops, defining scoring rules and running post-training for clients. That is higher-value work than basic chatbot builds.

US businesses, especially in legal, finance and healthcare, care most about control and accuracy. A model that learns their standards is appealing. But they will also ask hard questions about where their data goes and who checks the model.

Is AI Model Training Right for Your Next Project?

Here is the honest answer we give clients. Most projects don't need custom training yet. Training makes sense only when several things are true at once:

The task is narrow and repeated, like reviewing one type of contract or tagging support tickets.
You have real feedback from experts, such as edits and corrections, not just raw data.
You can score the output clearly, so you know if the new model is better.
Cost or speed is a real problem with your current hosted model.
The client has agreed in writing that their data can be used for training.

If two or three of these are missing, stay with prompting and retrieval. Spend the effort on better evaluation instead.

The competition is growing

Trajectory is not alone. TechCrunch names Thinking Machines Lab, the Mira Murati startup behind the Tinker fine-tuning tool, as a direct rival. Applied Compute, valued at $3.25 billion last month, and Prime Intellect are also in the space.

For agencies, more choice is good news. It also means you shouldn't lock a client into one platform too early.

A 6-Step Plan Before You Try AI Model Training

This is the checklist we use at Debugged Pro before we suggest training to any client:

Pick one task. Choose a single, high-volume job where mistakes are easy to spot.
Collect feedback first. Log every human correction for a few weeks before you train anything.
Write down what good looks like. Turn expert judgment into a clear scoring guide.
Get data consent. Update contracts so clients know how their data is used and stored.
Compare against a hosted model. Test the trained model on the same tasks and track cost, speed and accuracy.
Keep a human in the loop. Name a person who approves each new model version before it goes live.

The NIST AI Risk Management Framework is a helpful free guide for the risk side of these steps. If you'd like support, our software development consulting team helps clients plan and test AI builds like this.

Why the human role still matters

Trajectory's co-founder Michael Elabd made a point I agree with. He told TechCrunch that keeping humans involved in training makes the approach safer than fully autonomous self-improvement.

That links to a bigger debate we covered in our post on recursive self-improvement and Anthropic's IPO warning. Models that learn on the job need people who decide what they should learn.

Key Takeaways

Trajectory raised $45 million at a $300 million valuation, led by Sequoia.
Its platform retrains open source models on real feedback, an approach called continual learning.
Harvey, Rogo, Decagon, Mercor and Nvidia are named customers.
Performance claims come from the company and need your own testing.
Agencies should train models only for narrow tasks with good feedback, clear scoring and client consent.

Frequently Asked Questions

What is Trajectory and how much did it raise?

Trajectory is a startup founded about a year ago by three Stanford classmates. On November 2, 2026, it announced a $45 million Series A at a $300 million valuation, led by Sequoia, with Nvidia, Bessemer Venture Partners and Bain Capital Ventures joining.

What is continual learning in AI model training?

Continual learning means a model keeps improving after launch. It learns from feedback on real work, such as corrections and edits, and gets retrained on that signal. Most models today stay frozen until someone trains them again.

How is this different from fine-tuning?

Fine-tuning is usually a one-time project run by technical staff on a fixed dataset. Trajectory aims to make post-training ongoing and self-serve, so domain experts can define what good output looks like and the platform keeps retraining the model.

Should a small agency train its own AI model?

Usually not yet. Start with good prompts, retrieval and evaluation on a hosted model. Consider training only when you have a narrow, repeated task, enough quality feedback, a clear cost problem and written client consent to use their data.

Which companies already use Trajectory?

TechCrunch reports that Harvey and Rogo use the platform to post-train open source models, and that Decagon, Mercor and Nvidia are also customers. Trajectory says it is now making the platform publicly available.

About the Author

Nitin Kumar is Lead Project Coordinator at Debugged Pro. He manages web, mobile and AI delivery projects for clients in India and the USA, with a focus on practical, safe ways to use AI tools in real client work. You can find more of his articles on the Debugged Pro blog.