Alessandro Cappelli is an AI researcher and co-founder of Adaptive ML, the enterprise reinforcement-learning company acquired by Datadog in June 2026. He helped develop the open-source Falcon language models before building infrastructure that enables companies to train specialized models on operational feedback, evaluate their performance, and deploy them at scale.
At LightOn, Cappelli investigated alternatives to conventional neural-network computation. He led research on optical computing and adversarial robustness and co-authored a study of scaling laws beyond backpropagation, working alongside researchers who would later join him in founding Adaptive ML.
At the Technology Innovation Institute, he contributed to Falcon and its training-data foundations. He co-authored the RefinedWeb paper, which demonstrated the importance of filtering and deduplicating large web datasets, and the technical account of the Falcon model family.
In 2023, Cappelli co-founded Adaptive ML with Julien Launay, Daniel Hesslow, Baptiste Pannier, Axel Marmet, and Olivier Cruchant. Initially a research scientist, he later became chief customer officer. The company developed Adaptive Engine, combining model evaluation, reinforcement-learning-based training, and production serving for enterprises including AT&T, Manulife, and CCS.
- Production feedback as a training signal. Cappelli treats deployment as the beginning of continuous model improvement: customer interactions, business outcomes, evaluations, and operational failures become reward signals for subsequent training. Prompt changes and supervised fine-tuning, he argues, lack the same systematic feedback loop.
- Smaller models with workable enterprise economics. Reinforcement learning can specialize compact open models for particular tasks, lowering inference costs and latency while preserving control over sensitive data and model behavior. His work included adapting Gemma for multilingual content moderation with SK Telecom.
- Synthetic training trajectories for enterprise agents. Existing workflows can serve as training environments; when they are unavailable, simulated users and tools can substitute. Reward functions identify successful interactions, allowing rejection sampling to generate usable training examples. Real customer transcripts can help simulations reproduce difficult scenarios, including escalation to human agents.
- Human-defined rewards without perpetual annotation. People establish evaluation rubrics, tone requirements, escalation policies, and business objectives. Language-model judges apply those standards at scale, while accumulated production feedback supports increasingly specialized reward models.
In June 2026, Adaptive ML joined Datadog. Datadog acquired the company to advance world models and agentic post-training for observability.
