Ronak Malde's Trajectory raises $15M to ship continual-learning agents
Co-founders Michael Elabd and @QuantumArjun are building a platform that turns user corrections into post-training for deployed models; backers include Conviction and Bessemer.
By Ryan Merket ยท Published
Why it matters
If continual learning works safely in production, the advantage shifts from whoever trains the biggest model to whoever captures and uses in-product feedback loops best. That could turn usage into a compounding moat for AI-native apps.
Trajectory has launched with $15 million to build a continual learning platform, Ronak Malde (@rronak_) announced in a thread on X. Malde is co-founding Trajectory with Michael Elabd (@MichaelElabd) and @QuantumArjun, and says Trajectory will focus on making agentic AI systems get smarter from every user correction after they ship.
Malde framed the bet simply in his launch post: "AI will never again start on day one. Every correction, every retry, every edit will make products smarter. This is Continual Learning," he wrote on X. Digg noted that Malde has been predicting "2026 will be the year of continual learning" since January, and that the story was showing 98% positive sentiment and the top spot in its AI section at the time of posting, per Digg.
What Trajectory is building
In Malde's words, Trajectory is "a research lab and product company building the platform for Continual Learning." The pitch: unlock the signal already sitting in product usage so companies can continuously post-train large-scale agentic models after deployment. That means capturing retries, edits, and corrections as training signal so shipped models steadily improve without waiting for the next big pretraining cycle. Malde argues this can produce models that "outperform the frontier" by learning from real interactions rather than static datasets, per his launch post.
Who is backing it
Malde says Trajectory raised $15 million from @Conviction, @BessemerVP, @radicalvcfund, and individuals Jeff Dean (@jeffdean) and Fei-Fei Li (@drfeifei), among others. Digg reports that Conviction Capital led the round. Trajectory also named early partners such as @ClayRunHQ, @Harvey, @DecagonAI, @mercor_ai, and @RogoAI, with some already in production, according to Malde's post.
The founder's thesis and the status quo
The broader thesis challenges a common reality in AI product development: models remain static in production because user corrections rarely update them. Digg attributes that line of critique to Rohan Paul (@rohanpaul_ai) in its coverage, and positions Trajectory as an attempt to wire real-time product feedback directly into post-training pipelines. If continual learning can be made robust in production, the next leap in performance may come from how products learn after deployment rather than from pretraining alone, as Digg framed it in its writeup.
Malde also emphasized team composition, saying Trajectory has researchers from DeepMind, OpenAI, Apple, Meta Superintelligence, Amazon AGI, and Scale AI, alongside a product crew with experience at Stripe and Figma, per the launch post. Digg's thread adds that Malde previously led post-training at Windsurf before its Google/DeepMind acquisition, a detail it reported that is not otherwise described in the launch post.
What we do not know yet
Neither valuation nor round stage were disclosed. Trajectory has not shared technical details about its data pipelines, safeguards, or specific learning methods, and it has not broken out which named partners are fully live versus piloting. What is clear from Malde's posts is the bet: push learning to the edge of real product usage and let every customer correction make the model better.