Practical multimodal AI, from model research to on-premises deployment.
Roger Labs helps teams design, train and evaluate multimodal models, with a particular focus on time series forecasting. We also build the agentic workflows around those models and deploy the full stack on your own hardware when your data needs to stay in-house.
Roger Labs is led by Alexis Roger, a PhD candidate at McGill University and Mila - Quebec AI Institute, supervised by Professors Blake Richards and Irina Rish. His work centres on AI systems that are useful in the real world, especially when privacy, reliability and on-premises deployment matter.
Alexis has a background in mathematics and computer science, and has spent years designing, training and evaluating multimodal models. His research has run on the Frontier and Summit supercomputers, as well as through the AMD HPC Fund, and has led to papers at ICML, NeurIPS, and AAAI venues.
He has also worked with organisations where the margin for error is small: retrieval augmented generation and core banking customisations for major European banks, database access anomaly detection at Morgan Stanley, and patient-confidential clinical data at Stilla Technologies. Before that, he spent years hosting essential applications and network services for over 1500 users.
Research-grade AI systems that can be deployed, evaluated and maintained in production.
This is our core research area. We study time series foundation models at scale, including on some of the largest machines available, to understand what actually improves forecasting. That work points to a few practical lessons: language model pretraining can give forecasters a useful sequential prior, careful tokenisation can matter more than raw scale, and cross-modal transfer can often be done with small, reusable low-rank updates. We apply this work to demand, load and financial forecasting.
Vision-language models were our focus before time series, and much of the industry has since caught up. What matters now is what transferred: Robin, our released suite of multi-scale vision-language models, work on high-resolution image tiling, and methods for detecting and reducing hallucinations in long-form answers all came out of that period, and the lessons on aligning and evaluating a second modality are what we now bring to forecasting.
We build agents for real workflows, not just polished demos. That includes retrieval augmented generation over private document stores, research and engineering agents that keep results reproducible, tool orchestration across existing systems, and the guardrails and evaluations needed before a pipeline is left running.
Public leaderboards rarely answer the question a team actually has. We build evaluation suites around the task in front of us, including benchmarks like CHIRP for open-ended vision-language responses, robustness stress tests, data corruption checks and scoring methods that stay close to human judgement.
We support on-premises training and inference for organisations whose data cannot leave their own environment, from hardware sizing to the serving stack. The work is backed by hands-on hosting experience across DNS, DHCP, LDAP, VPN, storage, chat and web services for 1500+ users, plus privacy-conscious CCTV and surveillance deployments.
We work on adversarially robust models, ethics and morality evaluation for multimodal systems, privacy-preserving federated learning for intrusion detection, and anomaly detection for data access abuse. Our regulated-data experience spans banking, healthcare and finance.
Tell us what you are trying to build and what hardware you are working with. Choose a topic and send a message; nothing is submitted to a third party!