Company
One team, one core
Operators who have built and exited deep tech together, and the researchers who wrote RASP.
One core
Drawn as one system
Operators who have exited before, a research bench rooted at Mila, and advisors who built the field. Each wired to the core.
Fig. 1 · Org map · 2026 · Net: third venture together
The team
Founders, research, engineering, and operations
Built and exited three deep tech companies: Cilys (acq. Openwave), Neuralitic (acq. Guavus, then Thales), iPerceptions (acq. Emplifi). 26 years building systems that govern intelligence under finite resources.
Allocation, inspection, governance.
Former Research Scientist at Google DeepMind and NVIDIA. Led foundation model initiatives including pretraining 7B parameter LLMs. 2 ML patents, 10+ papers at ICLR, ICML, NeurIPS, EMNLP, ACL.
Built frontier models at DeepMind and NVIDIA. Now owns the layer that governs them.
VP Engineering at Hopper, Poka (acq. IFS), Plusgrade. Chief Software Architect at Neuralitic, VP Engineering at Guavus. Zero to eighty engineers, seed through Series E.
Four exits. Zero to eighty engineers, seed through Series E.
Core contributor on RASP and adaptive TEAL. Training free sparsity and adaptive inference.
Training free sparsity, the move that lets open models punch up.
PhD University of Copenhagen (DIKU), modular language modeling research with Mila. Inference efficiency research.
A Copenhagen PhD, and the inference efficiency research the runtime delivers.
Senior Principal Engineer at Broadcom, with deep work in cloud infrastructure, hardware acceleration, and low-latency AI RAN. Previously at VMware, Kaloom, Huawei, Opal-RT, and Rheinmetall. Fractional advisory with Prizmal on the compute and infrastructure architecture that sits beneath the control layer.
Senior Principal Engineer at Broadcom; fractional advisory on the infrastructure beneath the control layer.
Search infrastructure at Hopper, data engineering lead at Plusgrade, full stack medical imaging at Intelerad. Building the Prizmal Service core runtime and control plane.
25 years keeping production systems alive under load.
Founder and CEO of Title Collections, the governance platform for art and high-value assets. Earlier, CFO of Bläckfisk, Groupe Sélection's smart-building arm. At PSP Investments he underwrote data center portfolios and learned what compute is worth years before the AI boom made it obvious.
Has held the CEO seat, the CFO seat, and the investor's seat. Owns the numbers behind Prizmal's price.
Former Vantrix. Owns finance, legal, and the operating spine.
Runs finance, legal, and the operating spine.
The research bench writes the methods behind Six patent families.
Scientific advisory
The people who pressure test the science
Scientific Director of Mila, the world's largest academic deep learning research center. Trained under Yoshua Bengio and Geoffrey Hinton. Founded Google's Montréal AI research lab and spent nearly a decade across Google Brain and DeepMind before returning to lead Mila. His research has shaped how the field thinks about generative models, representation learning, and zero shot learning.
Hugo is also a vocal advocate for applying AI to environmental sustainability. That conviction is why his lens matters to Prizmal: every token routed to the right model, every computation avoided, is energy that does not get spent.
Why advise Prizmal
"The AI industry is scaling compute faster than it's learning to use it wisely. Prizmal is working on the right mandate: making inference efficient without compromising capability. That's not just a business opportunity, it's an environmental imperative."
Pressure tests: The research direction itself. Monthly sessions that challenge our hypotheses, and the bridge into the Mila network.
Plus three senior scientists from top tier global research centers. Names withheld pending publication consent.
Faculty researcher in a leading deep learning lab, trained under a foundational figure in information theory. Brings the information bottleneck lens on what neural networks keep and what they can afford to lose.
Pressure tests: RASP error bounds: what the residual tail can drop without losing reasoning.
PhD from a top Canadian ML institute, researcher at a global industrial AI lab. Focus on continual learning and modular, composable models that adapt without forgetting.
Pressure tests: The co adaptive loop: how the routing policy sharpens without drifting.
Researcher at a leading European institute. Lead author of ICML oral presentation work with a major silicon partner: up to 2.8x inference speedups across mismatched vocabularies.
Pressure tests: The draft and verifier control loop inside the Prizmal runtime.
One on ones, sessions that challenge the hypotheses
Mila network, structured bridge to researchers
Roadmap, situates the work against where AI and infrastructure are heading
Backers
Backed by TandemLaunch Ventures and i4 Capital
Careers and team inquiries: careers@prizmal.ai