Latest Research

Open research on frontier AI.

We publish our findings on efficient architectures, privacy-preserving inference, and bilingual evaluation. The AI community benefits when knowledge is shared.

Safety & Alignment

Safe-by-Design Alignment: Guardrails as First-Class Model Components

We show that embedding guardrails, refusal policies, and uncertainty calibration as architectural components — rather than post-hoc filters — improves robustness and reduces deceptive behaviour with no measurable capability loss.

May 6, 202610 min read

Model Architecture

Sparse Attention Mechanisms for 50% Inference Speedup

Apr 18, 20269 min read

Privacy & Security

Privacy-Preserving Inference Under PIPEDA

Mar 28, 202612 min read

Multilingual

QC-Bench: A Benchmark for Quebec French Language Models

Feb 14, 20268 min read

Efficiency

Loss-Aware INT4 Quantization for Edge Deployment

Jan 22, 202610 min read

Model Architecture

Constitutional Alignment in a Bilingual Setting

Dec 9, 202511 min read

Efficiency

Carbon Accounting for Hydroelectric LLM Training

Nov 17, 20257 min read