A fundamental shortcoming runs through virtually every major AI model on the market today, whether it originates from a leading lab or an open source project: these systems do not independently learn from their errors once deployed. Without additional training, they remain static, unable to adapt based on real-world performance.

For businesses looking to apply AI in specialized domains like legal services or financial analysis, the typical workaround involves fine-tuning a general-purpose model for their specific needs. That process is rarely straightforward. It usually demands either building out a dedicated internal research team or contracting with highly specialized engineers - often referred to in the industry as forward-deployed engineers - both of which carry significant costs and logistical complexity.

A startup called Trajectory is working to change that. Founded roughly a year ago by three former Stanford classmates with professional backgrounds at Google DeepMind and Apple, the company has built a platform designed to automate the model improvement process, enabling AI systems to learn from user feedback and grow progressively more capable at handling specialized tasks over time.

A $45 Million Bet on Automated AI Improvement

Trajectory, which currently employs 18 people, has announced the close of a $45 million Series A funding round at a reported valuation of $300 million. The round was led by Sequoia Capital, with additional investment from Nvidia, Bessemer Venture Partners, and Bain Capital Ventures.

Co-founder Ronak Malde described the current state of foundational AI models using an analogy: they behave like an exceptionally intelligent doctoral graduate who is perpetually starting a new position. The value that most organizations actually need, he explained, is closer to someone who has spent decades developing deep, applied expertise in a particular role. The gap between those two profiles is precisely what Trajectory is trying to bridge.

Continual Learning as a Core Strategy

The technical approach Trajectory uses is known as continual learning. Organizations that adopt the platform can begin with an open source AI model and integrate Trajectory's system directly into their existing applications. From there, the platform passively collects performance feedback as the model operates in the real world. That feedback is then used to retrain the model on an ongoing basis, gradually refining its performance within the specific domain where it is being used.

According to Malde, this design means that professionals such as lawyers, physicians, and financial specialists can guide model improvements themselves, with minimal technical involvement. Rather than waiting for an engineering team to intervene, these domain experts can effectively shape how their AI tools evolve - a shift that could meaningfully reduce both the time and cost associated with deploying specialized AI systems.

Several companies are already using Trajectory's platform in production environments. Legal technology firm Harvey and financial analysis platform Rogo are among those using it to post-train open source models, helping those systems develop the kind of task-specific knowledge that mirrors the experience of a seasoned attorney or analyst. In practice, this means models that can learn from repeated exposure to tasks like reviewing a specific category of legal agreement or applying a particular company valuation methodology. Other customers on the platform include Decagon, Mercor, and Nvidia. With this funding announcement, Trajectory is also making its post-training platform broadly available to the public for the first time.

Malde summarized the platform's value proposition clearly: the organizations using Trajectory get to define what quality output looks like within their own context, and that definition gets translated directly into measurable improvements in their model's performance.

Human Oversight Distinguishes the Approach from Autonomous Self-Improvement

Trajectory's continual learning methodology draws some surface-level comparisons to a concept called recursive self-improvement, which describes the theoretical ability of an AI system to autonomously enhance its own capabilities. However, co-founder Michael Elabd draws a clear distinction between the two approaches. In Trajectory's model, humans remain actively involved throughout the training loop. Elabd argues this is a critical safety feature - fully autonomous self-improvement, without human checkpoints, introduces the risk of unintended consequences that could be difficult or impossible to reverse.

A Competitive Field Taking Shape

Trajectory is not operating in an empty market. The company identifies several direct competitors working in the post-training and model fine-tuning space. Among them is Thinking Machines Lab, a venture founded by former OpenAI Chief Technology Officer Mira Murati, as well as Tinker, a platform focused specifically on fine-tuning. Applied Compute, which was valued at $3.25 billion as recently as last month, also operates in an adjacent space, as does Prime Intellect, which helps businesses design and train their own AI agents.

Despite the competition, Malde believes Trajectory's most durable advantage lies in the composition of its founding team and the broader talent it has assembled. Malde himself was part of a small cohort of researchers that Google DeepMind acquired through its purchase of Windsurf. Elabd also brings direct experience from DeepMind. The third co-founder, Arjun Karanam, spent time at Apple working on generative foundation models. Beyond the founders, the company has brought on researchers and product professionals who previously worked at OpenAI, Stripe, and Netflix.

Why Design Is Central to an AI Training Platform

For a company whose core product is deeply technical in nature, Trajectory places notable emphasis on user interface and design - a priority that might seem unusual at first glance. Karanam explains the reasoning: the startup has deliberately built an interface that makes post-training workflows accessible to users who do not have a machine learning background. The goal is to ensure that the domain experts who understand what good model output looks like - the lawyers, analysts, and clinicians who use these systems daily - can actively participate in improving them, without needing to understand the underlying engineering. That accessibility, in Trajectory's view, is not a secondary concern but a central part of what makes the platform viable for the organizations it serves.