Gaétan Forestier

DGA | AI Technical Project Engineer

French Defence Procurement Agency · France · April 2025 – July 2026

• On the customer side, the single technical point of contact for more than a year on a multi-million-euro NLP project extracting events from text with named entity recognition, built by a contractor for about ten core users and, in time, a few hundred more.
• Wrote the requirements, the acceptance tests and the full technical specification for the next phases of the contract under French public procurement rules; replayed the tests with the users at each demonstration and tested the system hands-on during final acceptance.
• Ran the workshops and led part of the technical negotiations with the contractor; often called in as the AI specialist in meetings between senior decision-makers and technical teams.

Safran Helicopter Engines | Data Scientist, Health Monitoring (Intern)

Tarnos, France · April – September 2024 · six-month internship

• Developed a Python regression model for Arriel 2E and 2D helicopter engines, estimating fuel flow from physical engine parameters.
• Integrated the model into Safran's industrial Python pipeline, wrote unit tests and deployed the changes through the existing GitLab CI/CD pre-production and production workflow.
• The model is used in production every day for Arriel 2E/2D health monitoring, on hundreds of helicopters, enabling fuel-flow-based analysis where direct measurements are unavailable.
• Conceived and developed a reference-normalised fuel-consumption indicator for engine condition monitoring; subsequently extended by the team for diagnostic analysis.
• Built an offline, rule-based diagnostic tool to flag early signs of fuel-system actuator anomalies in recorded engine data.

Sysnav | Computer Vision / R&D Intern

Vernon, France · April – August 2022 · five-month internship

• Developed an object-oriented Python/Tkinter R&D prototype to visualise and clean sparse visual-inertial SLAM point clouds and explore structural feature detection for indoor 2D mapping.

Open-weight LLM inference, evaluation and agent integration

Personal lab · 2026

• Compared quantisation and KV-cache settings on a Lenovo ThinkStation PGX (NVIDIA GB10, the DGX Spark platform), measuring generation speed and aggregate throughput under concurrent workloads. Explored trade-offs between memory use, context length and concurrency.
• Deployed open-weight models with vLLM and SGLang using Docker Compose, and connected self-hosted inference endpoints to agent workflows.
• Configured and ran an existing tool-use benchmark to compare model configurations.

Personal software development and self-hosted services

Personal projects · 2026

• Used Hermes Agent, Codex, Claude Code and MCP-connected tools for personal software development and remote Linux administration over SSH.
• Set up Docker Compose services, remote access and backups; checked service persistence and recovery after restarts.
• Built Grafana dashboards for lab-wide, per-host and service monitoring.

CSIRO Image2Biomass, Kaggle | Silver Medal, 45th of 3,805 teams

• Adapted a colleague-developed notebook for image-based biomass estimation and ran DINOv3 fine-tuning and submission experiments, including staged backbone unfreezing; tracked experiments with MLflow.