Replacing fine-tuning engineers: Trajectory raises $45 million to automate post-deployment model training

Enterprises that try to deploy AI for specialized tasks face a silent dilemma: foundational models do not learn from their mistakes automatically without subsequent training cycles, and adapting them for legal or financial work has so far required hiring internal research teams or contracting costly technical consultants known as field-deployment engineers. This expensive loop prompted Trajectory to raise $45 million in a seed round led by Sequoia, at a $300 million valuation, to automate the entire layer and shift tuning capability to non-technical specialists through direct work interfaces.

The startup, founded about a year ago by three Stanford alumni who previously worked at Google DeepMind and Apple, relies on a methodology it calls continuous learning; the company first links an open-source model to its applications to collect feedback on operational performance, then the platform reuses that data to retrain the model autonomously and improve its efficiency over time. Co-founder Ronak Maladi likens foundational models to a PhD-level AI student starting their first job every daywhile organizations seek the efficiency of an expert who has spent thirty years on the same taskswhich is the gap the platform aims to bridge by making training available to lawyers and doctors without software intermediaries

The platform is engaged in an intense race in the post-deployment training market; it ranks Machine Thinking Lab, founded by former OpenAI CTO Mira Morati, the $3.25 billion-valued company Applied Comput, and Prime Intellect among its direct competitors. The 18-person team, comprising researchers and product designers recruited from OpenAI, Stripe and Netflix, seeks to differentiate its offering by building an interface that makes human-expert supervision a core requirement; co-founder Michael Al-Abed says fully autonomous self-improvement carries unquantified risks, while keeping a human specialist in the training loop safely defines regulatory quality standards.

Companies such as Harvey and Roku already rely on the platform to train open models for reviewing qualitative contracts and evaluating firms according to specific financial methodologies, alongside clients such as Decagon, Mercur and Infidia, which also participated as investors in the round together with Seesaw Venture Partners and Bin Capital Ventures.Opening the platform to the public turns model tuning from a rare engineering task into a routine operational practicethereby reducing reliance on external technical consulting firms that previously monopolized complex customization processes

This shift directly impacts automation projects in Gulf and Middle-East markets, where banks, law firms and service companies grapple with a shortage of field-deployment engineers and globally high salaries; instead of waiting for specialized engineering teams to fine-tune open-source models for local languages and regulations, this approach puts the development tool in the hands of legal advisors and financial analysts themselves. The equation no longer concerns building larger models, but empowering local teams to train existing models and correct their drifts directly within the work environment.