Trajectory Raises $45 Million at $300 Million Valuation to Automate AI Model Post-Training

Whether developed by industry giants like OpenAI and Anthropic or built upon open-source alternatives, current foundational artificial intelligence models share a fundamental and critical limitation: they do not learn from their own mistakes without further training and human intervention.

For businesses attempting to adapt these general-purpose models for specialized and high-stakes fields such as law, medicine, and finance, this shortcoming presents a persistent hurdle. Organizations typically get around this limitation by fine-tuning general models to suit their specific operational requirements. However, that custom adaptation process frequently requires hiring expensive internal research teams or relying heavily on costly outside consultants, commonly referred to in the technology sector as forward-deployed engineers, or FDEs.

Trajectory, an eighteen-month-old startup founded by three Stanford classmates with professional backgrounds at Google DeepMind and Apple, is now stepping forward to automate that exact process. The company aims to allow AI models to learn continuously from feedback and steadily improve at specialized, domain-specific tasks without requiring an army of dedicated engineers.

The eighteen-person enterprise announced that it has successfully raised $45 million in a Series A funding round. The financing values the young startup at $300 million and was led by venture capital powerhouse Sequoia, with additional participation from prominent backers including Nvidia, Bessemer Venture Partners, and Bain Capital Ventures.

Discussing the core limitation that Trajectory was built to address, co-founder Ronak Malde compared current foundational models to an exceptionally bright PhD student who is nevertheless perpetually starting a brand-new job on day one. "But what the economy values is someone who worked on the job for thirty years," Malde noted in an interview.

The startup aims to solve this persistent problem through an established methodology known as continual learning. Businesses that choose to utilize Trajectory can start with an open-source AI model and connect the startup’s platform directly to their internal enterprise applications. This architecture allows the system to collect real-time feedback on how the model performs in daily operations. Trajectory then takes that performance information and uses it to retrain the model securely, helping it become progressively better at specialized tasks over time.

With Trajectory’s platform in place, professionals such as lawyers, doctors, and other domain specialists can improve underlying AI models nearly automatically and on their own terms, according to Malde. That capability is designed to make the deployment of specialized, highly functional AI systems significantly faster and considerably cheaper for businesses of all sizes.

Early traction suggests a strong market appetite for this approach. Companies like Harvey and Rogo are already utilizing Trajectory’s platform to post-train open-source models. This allows those models to perform more like experienced lawyers or veteran financial analysts by learning directly from previous tasks—such as reviewing a specific, nuanced type of legal contract or valuing a corporate entity using a distinct, proprietary methodology. Trajectory’s growing customer roster also includes notable names like Decagon, Mercor, and Nvidia. Following its latest funding announcement, the startup is officially making its post-training platform publicly available to a broader market.

The people utilizing Trajectory are the ones who ultimately decide "what good looks like for their organization and can turn that into improvements for their model," Malde explained, emphasizing the importance of keeping domain expertise at the center of the training loop.

Trajectory’s approach to continual learning shares certain conceptual similarities with recursive self-improvement, or RSI, which is the theoretical idea that advanced AI systems can autonomously improve their own internal capabilities over time. However, co-founder Michael Elabd argues that keeping humans actively involved throughout the training process makes Trajectory’s approach fundamentally safer. He points out that entirely autonomous self-improvement loops could potentially lead to unintended, unpredictable, or even dangerous outcomes in enterprise environments.

Trajectory is far from alone in identifying the massive market opportunity presented by automating post-training for enterprise clients. The startup considers Thinking Machines Lab—founded by former OpenAI chief technology officer Mira Murati and known as the company behind the fine-tuning platform Tinker—to be a direct competitor. Other rivals include Applied Compute, which secured a valuation of $3.25 billion, and Prime Intellect, a firm that assists companies in building and training their own customized AI agents.

Despite the crowded and competitive landscape, Malde believes that the startup’s single greatest competitive advantage is its exceptional team. Malde himself was among the handful of top-tier researchers that Google DeepMind acqui-hired from Windsurf. Elabd also brings valuable experience from his time at DeepMind, while Trajectory’s third co-founder, Arjun Karanam, previously worked on advanced generative foundation models at Apple. Furthermore, the startup has successfully recruited top-flight researchers and product designers from companies such as OpenAI, Stripe, and Netflix.

When asked why user experience and design matter so much for a deeply technical company focused primarily on machine learning training, Karanam explained that the startup has deliberately developed an intuitive interface designed to make the complex post-training process fully accessible to its nontechnical enterprise customers.