Machine Learning Research Engineer @ Windborne – Application Challenge
GitHub: https://github.com/Hrrsmjd
Which of our open ML Research Engineer roles are you interested in, in order?
General -> Applied Research -> Model Evaluation. The General role’s responsibilities are most closely aligned with my background and the work I enjoy, especially foundation models [1] and AI-based data assimilation [2].
Describe an ML idea you were initially excited about but later decided was wrong or unimportant. What changed your mind?
Early in my PhD, I was excited about developing novel architectures–often inspired by numerical methods or physics-based inductive biases. However, after many experiments, I have come to the conclusion that data and enabling new capabilities (and optimization) are more important. I shifted my focus to learning from corrupted measurements [2] and adaptive anytime-prediction models [3].
Describe a time when you embarked on a sidequest / took initiative outside your core job responsibilities. What value did you add to your organization? What drove you to action?
Two weeks before the 2024 NeurIPS Foundation Models for Science deadline, I started a side project to gain hands-on experience with foundation models. I preprocessed data, adapted Microsoft’s Aurora weather model, and trained the first foundation model of the Sun’s atmosphere. The paper [1] was accepted and led to a collaboration with NASA and IBM on a larger foundation model trained on 1024 GPUs.
How could better accuracy fail to translate into more useful forecasts for a weather model?
In many applications, better accuracy does not necessarily imply more useful forecasts. For example, infrastructure safety or insurance pricing may require being able to sample forecasts (including extreme events) from some distribution with calibrated uncertainty. Similarly, energy traders may require local, low-latency models with fine spatial and temporal resolution.
Describe a time you changed how you use LLMs in your work – switching models, tools, or workflows. What did you observe that prompted the change?
I initially saw little value in OpenClaw-like agents. In late 2025, I concluded that modern LLMs were good enough to monitor training and tune hyperparameters, which would free up my time for other projects. I built Hyppo, an agent that proposes and monitors experiments, then updates its prior and repeats the loop, while managing a compute budget. Today, I use agents to monitor my experiments.
(Optional) Your fav ML Twitter account?
@dwarkesh_sp (Dwarkesh Patel) for his technical discussions and lectures which help me keep up to date with LLM research.