The most expensive bottleneck in specialized AI isn't compute or data. It's the researcher who has to stop everything, collect a new dataset, and fine-tune the model again, because the lawyers found a new way the outputs were wrong. Trajectory, a San Francisco startup, is betting that bottleneck is a software problem. Their platform connects an open-source model to a live application, watches what users actually do with it, and uses those signals to retrain the model, continuously. The goal is to make the process so straightforward that a domain expert, not a machine learning PhD, can steer the improvement. It’s a bet that the next leap in AI utility won't come from bigger foundation models, but from models that learn on the job [Trajectory.ai].
Founder and CEO Ronak Malde frames it as the shift from static to living software. "Every product of the future will be a living system," he told Dutch Startup TV in June 2026 [dutchstartup.ai, June 2026]. The technical core is a method his team calls On-Policy Self-Distillation, or OPSD, which they say enables continual learning with less infrastructure and more granular, token-level feedback [SozAI]. For a company like Harvey, which builds AI for legal work, it means the model can learn from every time a lawyer edits a drafted clause or rejects a suggested precedent.
The wedge is accessibility, not just automation
Trajectory isn't the first to think about using production data. Large tech companies have internal systems for this. The startup's reported wedge is packaging the complex pipeline of data collection, evaluation, and retraining into a platform accessible to teams without dedicated ML engineering staff [TechCrunch, November 2026]. Their SDK requires what they say are just two code changes to start capturing traces and telemetry [Trajectory.ai]. The promise is that a product manager or a senior analyst could look at a dashboard of model performance, identify a failure mode, and trigger a retraining cycle aimed at that specific weakness, all without writing a line of training code.
This positions Trajectory not as a replacement for OpenAI or Anthropic, but as a layer on top of them, or more likely, on top of open-source models like Llama or Mistral. Companies start with a base model, connect it via Trajectory, and let real-world use shape it into a specialized tool. The business model is SaaS, with pricing likely scaling with model size and training frequency.
A team built on a $2.4 billion acquisition
The founders met their first week at Stanford and rebuilt their professional paths around advanced AI research [Jerry Chan - LinkedIn]. Their credibility is the engine here. Ronak Malde was an AI researcher at Windsurf, a startup focused on agentic coding, and followed the team to Google DeepMind after Google's reported $2.4 billion acquisition of Windsurf's top talent [Wired, May 2026]. He has said he left unvested Google equity on the table to start Trajectory [BigGo Finance]. Cofounder Michael Elabd worked in Google DeepMind’s robotics division, and Arjun Karanam worked on generative foundation models and the Vision Pro at Apple [Wired, May 2026]. It’s a team built at the intersection of ambitious research and product-scale engineering.
That pedigree translated into remarkable fundraising velocity. They raised a $15 million seed at a $115 million post-money valuation in May 2026, led by Conviction with Bessemer Venture Partners and others participating. Just six months later, in November 2026, they closed a $45 million Series A at a $300 million valuation, led by Sequoia with Nvidia, Bessemer, and Bain Capital Ventures joining [TechCrunch, November 2026]. The $60 million total war chest is for an 18-person team as of that November date, a signal of intense investor conviction in both the technical thesis and the team's ability to execute it.
| Round | Date | Amount | Lead Investor | Key Participants |
|---|---|---|---|---|
| Seed | May 2026 | $15M | Conviction | Bessemer Venture Partners, Radical VC, BoxGroup [Wired, May 2026] |
| Series A | November 2026 | $45M | Sequoia | Nvidia, Bessemer Venture Partners, Bain Capital Ventures [TechCrunch, November 2026] |
Early traction with AI-native builders
For a platform launched in November 2026, Trajectory has lined up a notable set of design partners and early customers, all of them startups building AI deeply into their core product. The named list includes Harvey (legal AI), Rogo (financial AI), Decagon, Mercor, and reportedly Nvidia [Yahoo Finance]. These are exactly the kind of customers Trajectory needs: companies where AI performance is the product, and where even incremental improvements in accuracy or relevance directly impact customer retention and revenue.
The deployment pattern, as described in coverage, involves connecting Trajectory's infrastructure to the customer's application. The platform then captures signals,user edits, corrections, retries, and rejections,and turns that raw telemetry into structured data for evaluating and retraining the underlying model [Built In]. For these partners, the value proposition is moving from periodic, costly fine-tuning projects to a continuous, automated learning loop.
The risks in the feedback loop
The bet is elegant, but the path is mined with technical and market risks. The platform's success hinges on a few critical assumptions holding true.
- Data quality over quantity. The raw signal from user interactions,a correction here, a re-prompt there,is noisy and implicit. The platform must reliably distinguish a user correcting a genuine model error from a user changing their mind or making a mistake themselves. Bad feedback could train a model to be confidently wrong.
- The compute cost of constant learning. Continually retraining even midsize models isn't free. The unit economics of Trajectory's service must allow customers to see a clear ROI from the incremental model improvements, after accounting for their cloud training bills. If the cost of learning outweighs the value of the improvement, the system stalls.
- The incumbent's move. Large cloud providers (AWS SageMaker, Google Vertex AI) and AI platforms (Weights & Biases) already offer tools for model monitoring and retraining. Their move into a more automated, continuous workflow is a matter of when, not if. Trajectory's defense is a superior developer experience and a focus on making the process accessible to non-experts, but they will be competing with integrated suites and vast sales teams.
The company's answer to these risks is its technical moat in OPSD and its early lead in productizing the workflow. The founder's research background suggests they are betting on staying ahead on the algorithmic curve, while the focus on design partners like Harvey is about proving the ROI in demanding, real-world environments before scaling.
The next twelve months
With the Series A closed and the self-service platform launched, the next phase is about moving from design partners to a broader customer base and proving the system works at scale. Key milestones to watch will be public case studies from initial partners showing measurable performance lifts, and an expansion of the partner list beyond the current circle of AI-native startups into more traditional enterprises with specialized AI needs.
The back-of-the-envelope calculation for a customer like a legal AI tool is straightforward. If a team of three ML engineers costing $600,000 annually is freed from quarterly fine-tuning cycles to work on new features, and the Trajectory subscription plus incremental compute costs come in at under $200,000 per year, the ROI is clear. The harder math is whether the continuous improvements drive enough additional user engagement or allow for price increases to cover the ongoing cost. That's the proof Trajectory needs to deliver.
For now, Trajectory is the most funded and credentialed startup trying to turn the academic concept of continual learning into an enterprise software category. The incumbent it must beat isn't another startup, it's the inertia of the old workflow: the scheduled, expensive, and disruptive fine-tuning project. If they can make learning from use as routine as pushing a software update, they won't just be selling a platform. They'll be defining how AI gets built.
Sources
- [Trajectory.ai] Trajectory homepage | https://www.trajectory.ai/
- [TechCrunch, November 2026] Trajectory raises $45M to make AI model training accessible to nontechnical experts | https://techcrunch.com/2026/11/02/trajectory-raises-45m-to-make-ai-model-training-accessible-to-nontechnical-experts/
- [Wired, May 2026] Former Google and Apple Researchers Launch a Startup... | https://www.wired.com/story/ex-google-apple-ai-researchers-want-to-make-ai-that-gets-smarter-as-you-use-it/
- [dutchstartup.ai, June 2026] Every product of the future will be a living system, Ronak Malde, Trajectory.ai | https://www.dutchstartup.ai/en/tv/every-product-of-the-future-will-be-a-living-system-ronak-malde-trajectory-ai
- [SozAI] Scaling up Continual Learning, Ronak Malde, Trajectory, Transcript & Summary | https://sozai.app/transcript/scaling-continual-learning-ronak-malde-trajectory/
- [Jerry Chan - LinkedIn] Jerry Chan LinkedIn post | https://www.linkedin.com/in/sagnak/
- [BigGo Finance] Ronak Malde Says Software Will Get Smarter Every Time You Use It | https://finance.biggo.com/news/dbf6d45da8ffc912
- [Built In] Trajectory company profile | https://builtin.com/
- [Yahoo Finance] Trajectory partners with AI-native companies | https://finance.yahoo.com/
- [StartupHub.ai, 2026] Trajectory's Ronak Malde on Scaling Continual Learning | https://www.startuphub.ai/ai-news/ai-research/2026/trajectory-s-ronak-malde-on-scaling-continual-learning