Ronak Malde

Ronak Malde’s AI work spans agentic coding and continual learning: improving systems from experience after deployment. He helped ship Windsurf, trained SWE-1 and contributed to the DeepMind team behind Antigravity and Gemini 3. After announcing his departure from DeepMind, he was identified as Trajectory’s CEO in the Delta Podcast’s May 28, 2026 episode.[3][2][4]

In May 2026, Malde was working on bringing continual learning to enterprises at Trajectory. The company’s approach treats users’ edits, retries, acceptances and corrections as training signals. Its premise is that actual user behavior reflects task success better than curated benchmarks. Trajectory describes a platform that captures these interactions and improves models, prompts and the surrounding agent harness, with model updates passing through customers’ evaluation suites and approval workflows before deployment.[2][5]

Malde contributed to writing and artifact creation for “Scaling SDPO”, alongside Arjun Karanam, Irene Han, Albert Li and Hersh Godse. The June 2, 2026 report investigates how to learn from individual agent trajectories that arrive after the model has already changed.[6]

The report builds on Self-Distillation Policy Optimization, introduced by Hübotter and collaborators. A model acts as its own teacher: the student sees the original task, while the teacher uses the same weights with additional information, such as feedback that became available afterward. Training brings the student’s token probabilities closer to the better-informed teacher’s. This supplies feedback throughout a single trajectory, including a failed attempt, without requiring a group of alternative responses to the same request.[6]

Stale trajectories complicate that recipe. The report describes training collapse with a naive importance-sampling correction, then uses PPO clipping to bound extreme weights and advantage clipping to limit individual token updates. On APEX-Agents, the team reported a 25% pass rate with GPT-OSS-120B, compared with a 5% zero-shot baseline and nine percentage points above SDPO without the stabilizers. Those experiments used one trajectory per task, retained failures and allowed three training steps of staleness. Learning from live production traces remained future work in the report.[6]

Before Trajectory, Malde worked on agentic coding at Codeium and Google DeepMind. In his May 24, 2026 departure announcement, he recalled joining Codeium a couple of years earlier, helping ship Windsurf and training SWE-1, which he described as a frontier agentic coding model. At DeepMind, he was part of the team that shipped Antigravity and contributed to Gemini 3.[4][2]

Malde credited Varun Mohan and Douglas Chen with showing him how a startup could move quickly and ship relentlessly, and Nicholas Moy with demonstrating effective research leadership. Announcing that he had left DeepMind, he also said he had decided to leave acquisition-related money behind. Four days later, the Delta Podcast’s May 28 episode identified him as Trajectory’s CEO, working on enterprise continual learning.[4][2]

His departure post also set out his research judgment: “It’s easy to optimize for the wrong benchmark or capability.” He advocated regularly reconsidering whether a team was pursuing the right direction. For speed, he emphasized concrete bets about where the world would be in six months, alignment around those bets and cutting other work. He illustrated that lesson with the sequence Codeium Extension → Windsurf IDE → SWE-1 → Antigravity → Antigravity CLI.[4]

In earlier coursework, Malde collaborated with Arjun Karanam on a reinforcement-learning project. Their CS 238 project trained a Rocket League bot specifically for kickoffs, testing curriculum learning and negative reward functions. The authors reported that curriculum learning did not improve results, while negative rewards produced kickoff metrics comparable to a human player and leading hard-coded bots. Their stated ambition to extend the approach to defending and attacking remained future work; the reported results concerned kickoffs rather than full-game play.[7]