Trajectory co-founders Ronak Malde, Michael Elabd and Arjun Karanam, whose startup makes AI model training easier — Trajectory's three co-founders. Image: Trajectory, via TechCrunch.

AI model training has always needed specialists. You hire researchers, or you pay expensive consultants to fine-tune a model for your field. A startup called Trajectory wants to change that, and investors just gave it $45 million to try.

The idea is simple to say and hard to build. Let lawyers, analysts and doctors improve a model themselves, using feedback from daily work. No research team in the middle.

I'm Nitin from the Debugged Pro team. Here is what was confirmed, how it works, and what agencies in India and the USA should do about it.

What Trajectory Announced

The news was first reported by TechCrunch reporter Marina Temkin on November 2, 2026. The Economic Times' Enterprise AI desk reported the same figures the next day. Here are the key facts:

The round: A $45 million Series A at a $300 million valuation.
Investors: Sequoia led, with Nvidia, Bessemer Venture Partners and Bain Capital Ventures joining.
The team: 18 people, founded about a year ago by three Stanford classmates.
Background: The founders worked at Google DeepMind and Apple.
Customers: Harvey, Rogo, Decagon, Mercor and Nvidia.
Availability: The post-training platform is now being made publicly available.

One small note from our own check. When we visited the Trajectory website, it still showed a "Join the waitlist" button. So access may roll out in stages.

The Problem It Wants to Solve

TechCrunch puts the core issue plainly. Today's AI models, whether from OpenAI, Anthropic or open source, don't learn from their mistakes without further training.

Businesses in law and finance usually fix this with fine-tuning. That often means building an internal research team. Or it means hiring consultants known as forward-deployed engineers, or FDEs.

Co-founder Ronak Malde used a sharp comparison. He said foundation models are like a very smart PhD student who's always 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 Changes AI Model Training

Trajectory's approach is called continual learning. The model does not freeze at launch. It keeps getting better as people use it.

Based on the TechCrunch report and the company's site, the flow looks like this:

Start with an open source model. The business picks a base model.
Connect it to your app. Trajectory says this takes a lightweight SDK and two code changes.
Capture real signals. Traces, corrections, edits and re-prompts from actual users.
Retrain on that feedback. The model improves at the specific task.
Approve before release. The company says updates pass your eval suite and approval workflow first.

Diagram of the continual learning loop used for AI model training, from open source model to approval — The continual learning loop, simplified. Sources: TechCrunch, Trajectory. Diagram: Debugged Pro.

Real Examples From Law and Finance

Harvey and Rogo already use the platform to post-train open source models. The goal is to make them act more like experienced lawyers or financial analysts.

The examples are specific. One is reviewing a particular type of legal contract. Another is valuing a company with a set method. That narrow focus is where this kind of training tends to pay off.

Malde said users decide "what good looks like for their organization and can turn that into improvements for their model." That line matters for agencies, and I'll come back to it.

Why Humans Stay in the Loop

Continual learning sounds a bit like recursive self-improvement, the idea of AI improving its own abilities. TechCrunch notes the overlap.

Co-founder Michael Elabd argues the difference is human control. Keeping people involved in training makes the approach safer, he says. Fully autonomous self-improvement could lead to unintended or even dangerous outcomes.

The Post-Training Market Is Getting Crowded

Trajectory is not alone. TechCrunch names several rivals going after automated post-training:

Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, which makes the fine-tuning platform Tinker.
Applied Compute, valued at $3.25 billion last month.
Prime Intellect, a platform that helps companies build and train their own AI agents.

Chart comparing companies working on automated AI model training and post-training — Companies chasing automated post-training. Source: TechCrunch. Chart: Debugged Pro.

When this many well-funded teams chase one problem, prices usually fall and tools get easier. For agencies, that is both a threat and an opening.

What This Means for AI Agencies in India and the USA

Let's start with the uncomfortable part. A lot of agency AI work today is custom fine-tuning and prompt tuning. If a lawyer can improve a model by clicking through corrections, some of that work goes away.

But look closely at what the platform still needs. Someone has to connect the app, capture clean signals, define good outputs, and own the release process. Those are not small jobs.

In our project work, the hard part is rarely the training run itself. It is agreeing with the client on what "correct" means. It is also getting usable feedback data out of a messy workflow.

Where Agencies Still Add Real Value

Integration. Wiring feedback capture into CRMs, ticketing tools and internal apps.
Evaluation design. Building test sets that reflect the client's real cases.
Data governance. Deciding which data can train a model, and which must never be used.
Deployment and monitoring. Rolling back a bad update quickly when needed.
Change management. Teaching domain experts how to give useful corrections.

For clients in India, cost is often the first question. Open source models with targeted post-training can be cheaper to run than large closed models for narrow tasks. For US clients, the first question is usually compliance and audit trails.

A Practical AI Model Training Plan for Your Agency

You don't need to buy any platform today. But you can prepare your team and your client projects now.

1. Start Capturing Feedback Now

Every AI product you ship should log user corrections, edits and re-prompts. Do this with consent and clear data rules.

Even if you never retrain a model, this data shows where the AI fails. If you do retrain later, it becomes your training set.

2. Build Small Evaluation Sets

Ask the client's best people for 50 to 100 real examples of good work. Turn those into a simple test you can rerun after every change.

This is the "value function" idea in practice. Without it, nobody can tell if a new model is better or just different.

3. Pick One Narrow, High-Volume Task

Don't try to train a general assistant. Pick something repetitive and well defined, like tagging support tickets or checking one contract clause.

Compare three options: better prompts, retrieval, and post-training an open source model. Go with whatever wins on your eval set at a sensible cost.

4. Agree on Data Ownership Early

Feedback data is valuable. Put in writing who owns it, where it is stored and whether it can train shared models.

This matters more under India's data protection rules and US state privacy laws. Get the client's legal team involved before launch, not after.

5. Keep Humans in Charge of Releases

Never let a retrained model go live without a human sign-off. Trajectory itself makes approval a core part of its pitch.

The NIST AI Risk Management Framework is a good, free guide for this. It helps teams map, measure and manage AI risks in a structured way.

6. Track Your Costs Per Task

Training runs, evaluations and hosting all cost money. Measure cost per completed task, not just monthly bills.

We wrote about this in our guide on the real cost of AI agents for agencies. The same habits apply to model training.

Should You Build or Buy?

This depends on the client. A simple rule we use:

Situation	Better fit
Narrow task, lots of daily feedback	Continual learning platform
Strict data rules, on-premise needs	Custom pipeline your team controls
Low volume, changing requirements	Better prompts and retrieval first
Early idea, no users yet	Ship an MVP and collect data

If a client needs a custom build, our LLM development services cover model selection and fine-tuning. For keeping models healthy after launch, MLOps services handle monitoring, versioning and safe rollouts.

What We Don't Know Yet

Some details were not in the reports we reviewed, so we won't guess:

Pricing for the public platform has not been shared.
There are no published benchmark results comparing it with rivals.
Regional availability, including for India, is not clear.

We'll keep watching. Funding rounds tell you where investors are betting, not what works in production.

Key Takeaways

Trajectory raised $45 million at a $300 million valuation, led by Sequoia.
Its goal is simple AI model training for experts like lawyers and analysts, using continual learning.
Humans stay in control, with evaluations and approvals before any update ships.
The market is crowded, with Thinking Machines Lab, Applied Compute and Prime Intellect in the race.
Agencies should prepare now by capturing feedback, building eval sets and settling data ownership.

Frequently Asked Questions

What is Trajectory and how much did it raise?

It is an 18-person AI startup founded by three Stanford classmates. It raised a $45 million Series A at a $300 million valuation.

What is continual learning in AI model training?

It means a model keeps improving after launch by learning from real feedback. Updates are retrained on corrections and edits, then approved by people.

Does this replace fine-tuning by AI engineers?

Not fully. It cuts some routine work, but integration, evaluation, data rules and deployment still need skilled teams.

Who already uses Trajectory?

TechCrunch names Harvey, Rogo, Decagon, Mercor and Nvidia as customers.

What should agencies do about this trend?

Capture feedback data, build small eval sets and agree on data ownership early. Then test post-training on one narrow task.

About the Author

Nitin Kumar is the Lead Project Coordinator at Debugged Pro. He works with AI and software project teams serving clients in India and the USA. He writes about practical ways agencies can use new AI tools without cutting corners.