# Trajectory Raised $45M for Continual Learning. Should You Care?

> Trajectory raised a $45M Series A to retrain open source models on customer feedback. The lesson: your feedback data and your definition of good work are the moat.

- Canonical: https://zeroraised.com/blog/trajectory-45m-continual-learning
- Published: 2026-11-03T09:24:47Z
- Author: ZeroRaised Team (https://zeroraised.com/blog/authors/zeroraised)
- Publisher: ZeroRaised (https://zeroraised.com/)
- Section: Growth
- Tags: artificial-intelligence, saas, bootstrapping, growth

Trajectory has raised a $45M Series A at a $300M valuation to make continual learning usable by people who are not machine learning researchers. For a bootstrapped founder, the round itself is not the story. The story is the asset underneath it: a structured loop around your feedback data and your definition of good work, which is something a small team can start building long before it can afford to train anything.

**TL;DR**

- Trajectory raised a $45M Series A at a $300M valuation led by Sequoia, with Nvidia, Bessemer Venture Partners and Bain Capital Ventures.
- It starts from an open source model, connects to your applications to collect feedback, then retrains the model on specialised tasks.
- Harvey and Rogo use it to post-train open source models, and other customers include Decagon, Mercor and Nvidia.
- Competitors include Thinking Machines Lab, Applied Compute and Prime Intellect.
- The moat is not the base model but the feedback loop and the explicit definition of what good looks like.

**Disclosure:** ZeroRaised publishes this analysis and is a launch site for bootstrapped products. Facts come from the linked sources as checked on 3 Nov 2026; ZeroRaised has no commercial relationship with the companies named.

**Continual learning** is an AI learning approach that trains a model sequentially on new tasks while preserving what it already learned, so it improves from a continuous stream of changing data instead of a fixed dataset ([IBM](https://www.ibm.com/think/topics/continual-learning)).

## What does Trajectory actually do?

The mechanics are simple to describe. A customer starts with an open source model, connects the platform to its applications to collect feedback on performance, and Trajectory retrains the model so it improves at specialised tasks over time ([TechCrunch](https://techcrunch.com/2026/11/02/trajectory-raises-45m-to-make-ai-model-training-accessible-to-nontechnical-experts/)). That is the whole loop: use, observe, retrain, repeat.

The company was founded about a year ago by three Stanford classmates, Ronak Malde, Michael Elabd and Arjun Karanam, who have experience at Google DeepMind and Apple. It has 18 people. The $45M Series A at a $300M valuation was led by Sequoia, with Nvidia, Bessemer Venture Partners and Bain Capital Ventures participating ([TechCrunch](https://techcrunch.com/2026/11/02/trajectory-raises-45m-to-make-ai-model-training-accessible-to-nontechnical-experts/)).

The customer examples are specific, which is useful. Harvey and Rogo use the platform to post-train open source models, for tasks such as reviewing a type of legal contract or valuing a company with a specific methodology. Other customers include Decagon, Mercor and Nvidia. The platform is now publicly available ([TechCrunch](https://techcrunch.com/2026/11/02/trajectory-raises-45m-to-make-ai-model-training-accessible-to-nontechnical-experts/)).

Malde's framing explains why the category exists. He compares foundation models to a very smart PhD student who is always starting a new job: "But what the economy values is someone who worked on the job for 30 years" ([TechCrunch](https://techcrunch.com/2026/11/02/trajectory-raises-45m-to-make-ai-model-training-accessible-to-nontechnical-experts/)). A general model arrives clever and inexperienced. The specialised model arrives trained on your work.

## Why is continual learning hard?

The core problem is catastrophic forgetting. A model trained on new data can lose previous knowledge, often because it overfits to the new data. The field describes the goal as balancing the stability-plasticity dilemma: learn the new thing without destroying the old thing ([IBM](https://www.ibm.com/think/topics/continual-learning)).

IBM groups the techniques into three families. Regularisation methods, such as elastic weight consolidation and learning without forgetting, constrain how much the model changes. Parameter isolation methods, such as progressive neural networks, give new tasks their own space. Replay methods keep a memory buffer of older samples, or generate them, and mix them back in during training. Replay needs storage, and it can be hard with sensitive personal data ([IBM](https://www.ibm.com/think/topics/continual-learning)).

IBM also separates three types: task-incremental, domain-incremental and class-incremental, with class-incremental described as one of the most difficult ([IBM](https://www.ibm.com/think/topics/continual-learning)). The labels matter less to a founder than the pattern: the more the world changes under the model, the more deliberately you have to preserve what it already knows.

Trajectory's answer is partly organisational. Users decide "what good looks like for their organization and can turn that into improvements for their model," Malde said. Elabd added that keeping humans involved makes the approach safer than autonomous self-improvement ([TechCrunch](https://techcrunch.com/2026/11/02/trajectory-raises-45m-to-make-ai-model-training-accessible-to-nontechnical-experts/)). That is a quiet but important design choice: the customer supplies the judgement, and the platform supplies the training machinery.

## Who is Trajectory competing with?

The market is already crowded. Competitors named by TechCrunch include Thinking Machines Lab, which offers the Tinker fine-tuning platform, Applied Compute, which was valued at $3.25 billion last month, and Prime Intellect. Trajectory's differentiation is an interface designed for nontechnical customers, according to Karanam ([TechCrunch](https://techcrunch.com/2026/11/02/trajectory-raises-45m-to-make-ai-model-training-accessible-to-nontechnical-experts/)).

That is a distribution claim as much as a product claim. If the tools are hard to use, the buyer is a research team. If they are easy to use, the buyer is an operations lead who owns a workflow and knows what a good output looks like. The second buyer is larger, but also less patient with vague results.

## What this means for bootstrapped founders

The strategic lesson is that your feedback data and your definition of quality are the durable assets. A competitor can rent the same base model. It cannot rent the two years of accept, reject and edit decisions your customers made, or the rubric your team wrote about what a good answer contains.

So start with the data, not the training. Log the input, the model output, whether a human accepted it, what they changed, and what happened next. Do this with consent and a clear retention policy. That record improves your prompts and evaluations immediately, and it is the raw material for any later training work. Our guides to [fine-tuning](https://zeroraised.com/blog/should-you-fine-tune-an-llm) and [evaluating LLM output](https://zeroraised.com/blog/how-to-evaluate-llm-outputs) cover the trade-offs.

Post-train when the task is narrow, repeated and expensive to get wrong, and when you already have enough labelled feedback to measure improvement. Do not post-train when the task changes every week, when you cannot yet evaluate quality, or when you have not tested cheaper options such as prompting and retrieval. If you are still choosing an approach, start with [what RAG is](https://zeroraised.com/blog/what-is-rag) and only move on when the evidence says you need to.

There is also a validation lesson here. Trajectory's model assumes the customer can define what good looks like in its domain. If you cannot write that definition down for your own product, training will not save you. Our guide to [validating a SaaS idea](https://zeroraised.com/blog/how-to-validate-a-saas-idea-before-building) is a useful check before you invest in a data pipeline.

Related reading from this week: [What Is Continual Learning? Why AI That Learns on the Job Matters](https://zeroraised.com/blog/what-is-continual-learning).

## Where ZeroRaised fits

ZeroRaised is a launch and discovery site for products built without institutional venture capital or accelerator equity. If you are building a product around feedback capture, evaluation or specialised model training, you can [submit your product](https://zeroraised.com/submit). The [pricing page](https://zeroraised.com/pricing) explains the current options. To follow the next batch of analysis, [join the ZeroRaised newsletter](https://zeroraised.com/newsletter).

## Sources

- [TechCrunch: Trajectory raises $45M to make AI model training accessible to nontechnical experts](https://techcrunch.com/2026/11/02/trajectory-raises-45m-to-make-ai-model-training-accessible-to-nontechnical-experts/)
- [IBM: What is continual learning?](https://www.ibm.com/think/topics/continual-learning)

## Frequently asked questions

### What does Trajectory do?

Trajectory runs a continual learning platform. A customer starts with an open source model, connects the platform to its applications to collect feedback on performance, and Trajectory retrains the model so it improves at specialised tasks over time.

### Is continual learning the same as fine-tuning?

They overlap but are not identical. Fine-tuning adapts a model with more data. Continual learning emphasises sequential training on new tasks while preserving what the model already learned, so it can keep improving from a changing stream of data rather than a single fixed dataset.

### When should a bootstrapped founder post-train a model?

Post-train when the task is narrow, repeated and expensive to get wrong, when you already have a meaningful set of labelled feedback, and when prompting, retrieval or ordinary fine-tuning have been tested and are not enough. Do not start with training if you cannot yet measure whether an answer is good.

### What is the cheapest first step?

Start logging feedback now: the input, the output, whether a human accepted or edited it, and the outcome. That record is useful for evaluation and prompting even if you never train a model, and it is the raw material you would need later.

### Is this relevant to a small SaaS product?

It can be, if the product performs a specialised task where customers can tell good output from bad. The value comes from the feedback and the definition of quality, not from the size of the model or the size of a funding round.