Kévin Osanlou

CV

Machine learning applied to planning, scheduling, combinatorial search and large language models. Full publication list under Publications.

Experience
2022 – present

AI Research Scientist & Consultant, Talan

Led applied research projects across graph machine learning and large language models: temporal graph networks for streamflow prediction with NASA collaborators, synthetic data generation with autoencoders, and LLM fine-tuning deployed behind FastAPI. Delivered production systems for national infrastructure operators, including retrieval and search over large technical document corpora, conversational assistants, and predictive models for network reliability.

2021 – 2022

Teaching Assistant, Paris-Dauphine

Natural language processing course.

2019 – 2020

Visiting Researcher, NASA Ames Research Center

Dynamic scheduling under uncertainty for UAV mission scenarios. Introduced a restricted form of dynamic controllability for disjunctive temporal networks, with a tree search guided by a graph neural network. Published at AAAI 2022 with NASA co-authors.

2017 – 2021

Doctoral Researcher, Safran Electronics & Defense

Industrial doctoral research on path planning for autonomous ground vehicles: graph neural networks used as heuristics inside exact solvers (constraint programming, branch and bound, A*, local search). Published at IROS 2019, with earlier results at the ICML and ICAPS planning workshops. Sole inventor on the resulting patent, granted in Europe and the United States.

Education
2017 – 2021PhD, graph machine learning and automated planning and schedulingParis-Dauphine / PSL
2014 – 2016MSc, Computer Science, Artificial Intelligence, with a machine learning specialisation (double-degree agreement with CY Tech)Paris-Dauphine
2013 – 2016Diplôme d'ingénieur (Master's-level), Computer Science, Data ScienceCY Tech (EISTI at the time)
Service
2022 – presentProgram committee reviewerAAAI
Technical
Machine learningPython, PyTorch, PyTorch Geometric, scikit-learn · fine-tuning (LoRA / QLoRA), retrieval-augmented generation, multi-agent pipelines
OptimisationConstraint programming, branch and bound, A* and local search · CPLEX (docplex) · C++ solver implementation
PlatformsAWS (Bedrock, Lambda), Databricks, FastAPI, CUDA, GPU training on rented instances