Gary KLAJER

Advisor Pierre-Antoine Bannier
Medical Imaging for histo-interpretability.

- Florence-2 Deployment @ OOrion
Embedded Florence-2 on iPhone for real-time object detection (TTFT 400ms, TPOT 20ms) for visually impaired

- French Lipreading @ La Ligue Contre Le Cancer
Built a video processing pipeline that ultimately gathered a high-quality, French, lip-focused annotated dataset (+500h) and benchmarked and trained SOTA open-source English Audio-Visual Speech Recognition models for a post-laryngectomy voice-preserving app.

- Software Automation @ Biogoup
Automated FSE processing for Biogroup labs by integrating browser (Playwright) and desktop (RobotJS) automation into a NestJS backend with React frontend, enabling secure CPS card authentication and streamlined teletransmission via Kalisil and Pyxbio under strict constraints.

Advisors J. Atif and O. Cappé
During my research internship with the MILES team, I investigated privacy vulnerabilities in language models applied to structured medical data. Using Medinote—a fine-tuned version of Meditron trained on JSON medical records—I introduced a novel targeted extraction (linkage) attack that leverages auxiliary public data to recover sensitive patient information. I designed and evaluated conditional sampling strategies, including prefix conditioning, and analyzed sampling schemes aimed at maximizing joint probability over full sequences. I also applied sequence-level perplexity metrics to perform Membership Inference Attacks, enabling more accurate discrimination between in- and out-of-distribution data to improve the filtering of non-member candidate samples.

Alongside my research, I conducted an extensive literature review on medical LLMs, privacy attacks (extraction, membership inference), DP techniques, unlearning, ICL, and advanced transformer variants (e.g., KNN-LLM, MoE, SSMs). I deepened my understanding by re-implementing models like bigrams, BPE, decoder-only transformers, and MoE in PyTorch, and by revisiting theoretical results on benign overfitting. I also fast-tracked my training in Differential Privacy through coursework (M2 IASD) and Dwork’s materials, and participated in reading groups on advanced topics such as higher-order optimization and optimal transport.

Helped students build strong foundations and deepen conceptual understanding through structured exercises, step-by-step explanations, and advanced problem-solving, fostering autonomy and broader insight into mathematical methods.