Jason Mancuso

Jason Mancuso is a machine-learning engineer who helped bring secure computation to TensorFlow and later worked with Modal customers on language models and agents. At Modal, he and his colleagues showed that the GPUs running an agent through a task could sit in different clouds from the cluster training the model. That let the team use less expensive hardware for those attempts, which account for much of the compute in agent training.

Computing where the data lives

Jason worked as a data scientist and contributed to bioinformatics research at the Cleveland Clinic, where he helped investigate whether blood test results could flag a contaminated sample. He also founded Cleveland AI. By 2019 he was a research scientist at Dropout Labs, which later became Cape Privacy. His open-source work in privacy-preserving machine learning included developing and maintaining TF Encrypted, PySyft, and Moose.

Jason and his co-authors showed in 2018 how to run secure multiparty computation inside TensorFlow. Several parties could compute a result together while keeping their individual inputs private. Implementing the protocol in TensorFlow meant researchers could express a model with familiar tools and use the framework’s existing optimizations for distributed computation. Customers had a related problem: they had permission to train on data but could not take possession of it. PySyft could send a computation to the machine holding the data. Jason and his colleagues extended that capability to TensorFlow, bringing it to developers beyond PySyft’s original PyTorch users.

Medical imaging gave those privacy requirements a concrete form. Jason co-authored the 2020 PriMIA study, in which three machines trained a shared classifier for children’s chest X-rays while each kept its images local. The researchers also demonstrated how a clinician could use a trained model as a service. The clinician sent an encrypted image and could decrypt the prediction, while the service kept its model weights private.

What an agent learns from the score it gets

In 2019, Jason, Tomasz Kisielewski, David Lindner, and Alok Singh studied corrupted rewards. In a grid, a square that reports an anomalously high reward can draw an agent away from the best route. The researchers identified conditions under which those spikes are distinguishable from genuine rewards. In two grid environments, their algorithm detected the misleading squares, and the agent still learned an optimal route.

In 2025, Jason was weighing which research direction to pursue. He credited conversations with Modal engineers about the systems they built for customers with his decision to learn more about language models and agents through hands-on customer work. He joined Modal as a forward-deployed machine-learning engineer.

Jason argues that teams have a reason to train their own agent when a general-purpose model struggles to act on proprietary data or when serving the model themselves would help them meet cost and latency targets. His starting point is a failed production session. A recording captures just one path through a task. With the agent’s usual tools and a test for success, the team can turn the failure into a task the model can attempt again, exploring other paths and learning from their outcomes. He develops this argument in Training Your Own Agent Models With RL Is Now Practical.

At Modal, waiting for a single cluster large enough to run all those attempts could leave available GPUs elsewhere unused. Jason and his colleagues kept training on a tightly connected cluster and sent workers in other clouds the compressed changes to the model’s weights after each update. Running attempts while training continued kept GPUs busy, but some attempts came from an older copy of the model. Different GPU types could also assign slightly different probabilities to the same token. The team corrected for those differences during training. In their experiment, training remained stable across the mixed fleet, and generating the attempts cost 15% less than in the single-GPU-type comparison.