Dario Fumarola
Forward Deployed Engineer, AWS
AI Researcher
I build AI systems with customers, from early experiments to deployment, and research how agents learn and adapt. Based in New York.
About Me
With AWS customers, I take a working experiment through deployment. Latency, integration and reliability matter as much as model quality, and I stay involved as the system goes into use.
At Amazon Science, I work on reinforcement learning and adaptive inference. I study how agents reason and cooperate, and how their behavior holds up when conditions change after training.
For research or engineering collaborations, get in touch.
Selected Work
Mood Swings: Calibrated Reward-Cost Control ICML 2025
Make a trained agent more or less cautious at deployment, without retraining.
98% of an eight-policy library's reward-cost coverage, at one-eighth the training.
Who Needs Attention Anyway? NeurIPS 2025
A controller remembers a planner's corrections so the next decision can reuse them.
About 23% fewer repeated derivative evaluations over the following eight decisions.
Broadcast-Gain: Coordinating Agents with Small Messages NY-RL 2025
Let agents report what they need so a shared controller can coordinate their next moves.
Two-byte demand reports guide junction access. A deadline keeps inconclusive reports from postponing a turn.
Research & Engineering Interests
Agent training
Reinforcement learning and model training for agents that reason, use tools and cooperate. I'm interested in what interaction teaches, and what kinds of experience an agent needs.
Memory & computation
A model can spend more compute on a hard decision or carry a useful correction into the next one. I study how those choices change its behavior after training.
Evaluation & deployment
Customer work starts with a task someone needs to get done. I build evaluations around that task and follow the system through integration, inference tuning and failure recovery.
Experience
Amazon Web Services
Forward Deployed Engineer · 2023 to present
Technical lead for more than 20 enterprise AI engagements, from customer discovery through production. For pharmaceutical research teams, I built deployment agents that automate model packaging, GPU scheduling, validation and recovery, cutting deployment time from weeks to hours.
Other work spans evaluation frameworks, Bedrock inference optimization and camera-to-radar data synthesis.
Amazon Science
AI Researcher · 2023 to present
I co-lead a Build on Trainium project studying whether training LLM agents through social interactions improves cooperation, reasoning and safety. I work on the training infrastructure and evaluation across checkpoints.
Education
Washington and Lee University
Bachelor's degree in Computer Science and Mathematics · 2019–2023
Undergraduate research on numerical methods for nonlinear differential equations.
Last updated: October 2026