Praedon, a deep-tech startup co-founded by HEC Paris Professor Maximilian Schiffer and Professor Axel Parmentier from École des Ponts ParisTech, has been selected for the Bpifrance i-Lab innovation grant. This prestigious award, amounting to the maximum funding of €600,000, was officially presented on September 16.
Praedon develops novel AI systems based on a "two-brain" architecture and Decision-Focused Learning (DFL). This technology aims to automate complex decisions in critical industrial environments where traditional generative AI models pose significant operational risks.
Unlike large language models (LLMs) that can "hallucinate" with minor consequences, Praedon's AI is designed to guarantee mathematically feasible decisions by integrating physical, legal, and technical constraints. It has been successfully applied in areas like energy trading, ensuring transactions adhere to physical infrastructure limits.
The Praedon platform functions as a backend decision engine, capable of processing uncertain variables such as weather events, volatile markets, or logistical bottlenecks to prescribe optimal, resilient decisions. Studies, including one co-authored by Professor Schiffer, demonstrate significant reductions in response times for emergency services, such as ambulance dispatching.
Transitioning from an academic breakthrough project, Praedon has become a scaling commercial entity. The technology was validated through consulting missions and industrial pilot projects, notably with Air France (achieving a 27% reduction in delay-induced costs) and Renault Group (generating over €1 million in annual savings and a daily 10% CO2 emission cut across Europe). Officially established in 2026, the company now offers its platform as a Software-as-a-Service (SaaS) and manages algorithmic energy trading through a joint venture with Flexbid.
The co-founders emphasize that their solution, by optimizing complex decisions, generates millions of euros in operational savings for industry leaders. Furthermore, due to its use of structured models, it requires significantly less computing power and energy than standard LLMs.
