Nolinor Aviation cuts safety-investigation time 80% with Mila LLM and agentic AI

Title

Nolinor Aviation cuts safety-investigation time 80% with Mila LLM and agentic AI

Content

Nolinor Aviation, a Quebec-based charter airline, partnered with Mila in late 2023 and reduced the human effort required for safety reporting and investigation by up to 80%, cutting the human time for a full investigation report from 40 hours to 5 hours while preserving reliability and accountability. “It is meant to be interactive — it’s not meant for the LLM to do it in one shot and then it’s done. It’s meant to allow back-and-forth, to be really interactive. In the end, the human has the responsibility of the report,” said Hadrien Bertrand, Senior Applied Research Scientist at Mila. Olivier Richer, Safety & Quality Management System Director for Nolinor, added: “This project is a perfect example of Quebec’s innovation expertise, where Canadian and Quebec funds are invested locally to support local companies. The result? Talented researchers from diverse backgrounds collaborate with specialists from the Canadian airline industry to spread the knowledge emerging from Mila’s ecosystem.” Aviation safety management has long been a meticulously human affair: when an employee witnesses a safety event, they file a report ranging from a few sentences to a detailed paragraph. Human investigators then compile those reports, conduct risk assessments, and decide whether to open a full investigation — a process that can absorb dozens of hours per event. Recognizing the potential to streamline that work, Nolinor partnered with Mila in late 2023 to automate large portions of it. The build played out in two phases. In phase one, Mila’s applied research team used a Large Language Model to process the variable-length employee safety reports and structure them into a standardized template — summary, sequence of events, event type, time, location, personnel involved, attachments. The model then performed a risk assessment using Nolinor’s internal methodology, identifying failed prevention and recovery barriers and evaluating scenario probabilities; outputs fed Nolinor’s risk matrix, with the LLM providing its reasoning before issuing a final risk score. In phase two, the team layered agentic AI on top: the LLM extracted timestamped event sequences and immediate corrective actions, while the agent queried aircraft and equipment manuals, flight information, weather data, and Nolinor’s internal database for employee training records. When the agent hit a roadblock, it generated a query for the human investigator — keeping a tight human-in-the-loop. The operational result is large. A full investigation report that previously required 40 hours of manual work is now done in five hours of human involvement — an 80% reduction. Investigators no longer spend time on tasks with little added human value; they focus on validation, verification, improvement of reports, and effective mitigation actions. As Richer noted, that efficiency matters at a time when skilled aviation safety workers are hard to hire. Recognizing the broader potential, Nolinor is launching a spin-off called CIRRUS Intelligence to commercialize the solution for the wider aviation industry — taking the Mila-Nolinor collaboration beyond one airline’s safety department.