Flemming Kondrup

Artificial Intelligence Researcher

Hi! I am a PhD Student at Mila and McGill University supervised by Doina Precup and Lars Grant and advised by Joelle Pineau. I previously completed a BSc at McGill University, for which I received the First-Class Honours distinction and the Dean’s Multidisciplinary Undergraduate Research List Distinction, during which I completed 5 research internships working with Doina Precup, David Juncker, Gabriel Venne and Peter Metrakos.

My main research interests are:

Enhancing the safety, reasoning, and sample efficiency of autonomous agents in complex environments by integrating LLMs and VLMs, with a focus on robust decision-making and generalization in web-based settings.
Developing agentic systems for healthcare that prioritize safety, robustness under uncertainty, and efficient learning in data-constrained environments, with a focus on reliable decision-making and real-world applicability.

I am a recipient of the:

Work Experience:

Machine Learning Intern – ServiceNow, Montreal, Canada

Developing a safety-aware autonomous web agent using RLHF to balance task success and adversarial robustness

Summer 2025 | Mentor: Gabriel Huang

Machine Learning Intern – Dialogue, Montreal, Canada

Led the successful deployment of Generative Vision-Language Models (VLMs) in production to automate patient photo verification for telemedicine, boosting classification accuracy by 17% and streamlining the intake process, and developed a novel LLM-powered symptom intake system to reduce patient input time and improve triage.

Jan – Apr 2025 | Mentor: Alexis Smirnov

Selected Research

Improving agent decision-making in complex environments:

Cracking the Code of Action: A Generative Approach to Affordances for Reinforcement Learning

Leveraging VLMs to guide RL agents and improve decision making in high-dimensional action spaces

Paper

Forecaster: Temporally Abstract Tree-Search Planning from Pixels

Hierarchical RL with abstract world models for tree-search planning

Paper

AI for Healthcare:

Transferrable Model-Based RL for Personalized Insulin Therapy

Combine LSTM Forecasting & RL for individualized treatment

Prospective Publication

Safe Mechanical Ventilation Using Deep Conservative Q-Learning

DeepVent, an offline AI agent for safely optimizing ventilator settings

Paper

Service and Leadership

In 2022, I was captain of McGill’s team in Project X, an AI research competition organized by the University of Toronto with competitors from top academic institutions across North America with renowned sponsors (Google, IBM, Moderna etc.). Our work on Deep Conservative Reinforcement Learning for Mechanical Ventilation received the highest score out of all 25 papers submitted, winning the competition with a $25,000 award, and leading to press interviews with The Tribune and The McGill Reporter.

In 2023-2024, I served as the Executive Director of the McGill Student Emergency Response Team (MSERT), overseeing a team of over 70 responders dedicated to providing emergency medical aid. In addition to managing a $100,000 budget and supervising a 7-member executive board, I facilitated communication between MSERT, the McGill University administration, and governmental agencies. My approach emphasized thoughtful leadership and fostering a collaborative team dynamic, enabling MSERT to expand its services and educational outreach. Over the past five years, I have also volunteered as a responder, contributing more than 2000 hours to the team’s efforts.