Fine-tuning a general model for a law firm or a bank usually means hiring research engineers or paying consultants who fly in to do it. Trajectory thinks the domain specialists should just be able to teach the model themselves.
What the platform does
Customers start with an open-weight model, wire the platform into their own applications, and let it collect feedback on how the model performs. Trajectory retrains on that signal, so the system improves at narrow tasks over time. Co-founder Ronak Malde frames the gap simply: foundation models are brilliant graduates who start a new job every morning, while the economy pays for thirty years of experience.
The round
Sequoia led the $45M Series A, valuing the 18-person company at $300M, with Nvidia, Bessemer Venture Partners and Bain Capital Ventures participating. The platform, previously in private use, is now publicly available. Early customers include Harvey, Rogo, Decagon, Mercor and Nvidia itself, mostly post-training models for contract review and valuation work.
Why it is a crowded corner
Thinking Machines Lab, Applied Compute and Prime Intellect are all chasing the same post-training budget, and Applied Compute was valued at $3.25B last month. Trajectory’s bet is that design, not just research pedigree, decides the winner. Co-founder Arjun Karanam argues the interface is what makes continual learning usable by doctors and lawyers rather than machine-learning engineers.
The team also draws a line against recursive self-improvement, the idea of models improving without oversight. Co-founder Michael Elabd argues keeping humans in the loop makes the approach safer, a distinction worth watching as AI labs tout ever more autonomous training loops.
