Trajectory raises $45M to automate AI model post-training

Trajectory has secured a $45 million Series A round at a $300 million valuation, led by Sequoia with participation from Nvidia, Bessemer Venture Partners, and Bain Capital Ventures. According to TechCrunch, the 18-person startup aims to solve a critical limitation in current AI deployment: the inability of models to learn from mistakes without extensive, expensive retraining. By automating the post-training process, Trajectory seeks to make specialized AI improvement accessible to domain experts in fields like law and finance, rather than restricting it to internal research teams or costly forward-deployed engineers.

The Continual Learning Approach

Co-founder Ronak Malde describes the problem using a specific analogy: foundational models are like a "very smart PhD student who’s always starting a new job," whereas the economy values the experience of someone who has worked on the job for 30 years. Trajectory’s solution involves connecting its platform to a company’s applications to collect feedback on model performance. This data is then used to retrain the model, allowing it to improve at specialized tasks over time.

The platform is built for users who can define "what good looks like for their organization" and translate those criteria into model improvements. Malde notes that this capability allows lawyers, doctors, and other specialists to improve models "near automatically," reducing the barrier to entry for deploying specialized AI. The company has now made its post-training platform publicly available, moving beyond its initial private beta phase.

Current Customers and Use Cases

Trajectory is already being used by several high-profile companies to post-train open-source models. Harvey and Rogo utilize the platform to help models perform more like experienced lawyers or financial analysts. Specifically, the system helps models learn from previous tasks, such as reviewing particular types of legal contracts or valuing companies using specific methodologies. Other customers include Decagon, Mercor, and Nvidia. By leveraging open-source base models and applying Trajectory’s continual learning loop, these companies aim to achieve performance levels comparable to fine-tuned proprietary models without the associated costs and complexity.

Competitive Landscape and Safety Considerations

Trajectory enters a crowded market where several startups are attempting to automate post-training. Competitors include Thinking Machines Lab, founded by former OpenAI CTO Mira Murati; Applied Compute, which was valued at $3.25 billion last month; and Prime Intellect, which focuses on building and training AI agents. While Trajectory’s approach shares similarities with recursive self-improvement (RSI), co-founder Michael Elabd argues that keeping humans involved in the training process makes the system safer. Elabd contends that autonomous self-improvement could lead to unintended or dangerous outcomes, whereas Trajectory’s human-in-the-loop design mitigates these risks.

The startup’s team composition is a key differentiator, according to Malde. The three co-founders are Stanford classmates with significant experience at major tech companies: Malde was among the researchers Google DeepMind acqui-hired from Windsurf, Elabd also worked at DeepMind, and Arjun Karanam previously worked on generative foundation models at Apple. The team has also recruited researchers and product designers from OpenAI, Stripe, and Netflix. Karanam emphasizes the importance of design in making AI training accessible, noting that the startup has developed an interface specifically tailored for non-technical customers.

Implications for Developers

For developers and technical leads, Trajectory’s platform offers an alternative to traditional fine-tuning pipelines. Instead of manually curating datasets and managing training runs, teams can integrate Trajectory into their application stack to capture real-world feedback. This is particularly relevant for organizations using open-source models, as it provides a structured way to improve performance on niche tasks without building a full-scale training infrastructure. However, developers should be cautious about the quality of the feedback data collected, as the effectiveness of continual learning depends heavily on the accuracy and relevance of the signals provided by the system. The platform’s public availability allows for immediate testing, but organizations should evaluate whether their specific use case aligns with the types of tasks where Trajectory has demonstrated success, such as legal contract review or financial valuation.

What to Watch

• Adoption Rate: Whether the public release leads to significant adoption beyond its current enterprise customers like Harvey and Nvidia.
• Competitive Response: How Thinking Machines Lab and Applied Compute respond to Trajectory’s $300 million valuation and public platform launch.
• Safety Audits: Independent verification of the safety claims regarding human-in-the-loop continual learning versus autonomous self-improvement.
• Model Compatibility: Which open-source models Trajectory supports best, and whether it works effectively with newer releases from major labs.