Stefano Fiorucci is an AI and software engineer at deepset and a contributor to Haystack, the open-source framework for building applications with language models. His work spans retrieval, multimodal systems, and specialized small-model post-training, with a particular focus on teaching compact models to reason through carefully engineered reinforcement-learning environments.
Earlier in his career, Fiorucci worked at 01S on extracting and retrieving information from unstructured documents. A fellow in the tenth session of the Pi School of AI, he later explored Italian-language model adaptation using supervised fine-tuning, preference tuning, and evaluation. His independent projects have also included retrieval-based fact-checking and multilingual synthetic instruction data.
At deepset, he co-led multimodal Haystack development with Sebastian Husch Lee, extending image support across model providers, prompt templates, document conversion, embeddings, and retrieval. He also built a browser-operating agent combining Haystack, Gemini, Playwright, and the Model Context Protocol.
In 2025, his Qwen Scheduler GRPO project used reward functions to train a model to prioritize events and assemble schedules. Its seven-billion-parameter model outperformed both its unmodified counterpart and a larger model on the project’s test set, although overlapping events exposed weaknesses in the reward design.
- Open reinforcement-learning environments. Fiorucci treats environments as reusable software combining task data, interaction rules, tools, state, and scoring. He advocates interoperable training environments so open models can develop capabilities without depending on fragmented or proprietary infrastructure.
- Small models can learn beyond their teachers. His LLM RL Environments Lil Course trains a small model to play tic-tac-toe through supervised warm-up followed by reinforcement learning with verifiable rewards. Initial examples come from GPT-5 mini; against an optimal opponent, the resulting specialist exceeds its teacher in Fiorucci’s task-specific evaluation.
- Evaluation must expose environmental shortcuts. Fiorucci uses minimax-based opponents, deterministic seeding, and stratified sampling to control difficulty and reduce noisy comparisons. He identified how predictable opponent behavior encouraged memorization, how small batches destabilized training, and how exclusively perfect opponents produced excessively defensive play. His remedy is direct: inspect trajectories, vary opponents, and test the trained model in the actual task.
His next focus is training compact models around small sets of practical tools and measuring whether focused post-training can outperform larger general-purpose systems on specific workflows.
