How Albatross Retrains Recommendation Models Multiple Times Per Day on GPU Infrastructure Tailored to Their Stack

A custom integration built by Lyceum's engineers and shipped in days lets Albatross iterate faster on recommendation models and only pay for compute they actually use.

AT A GLANCE
Client	Albatross
Industry	AI / Product Discovery
Use Case	Frequent model retraining, production inference, experimental R&D
Lyceum Products	Serverless GPU, Custom Integration
Key Result	Pay-per-use GPU economics, multi-daily model retraining shipped in days, with no changes to Albatross's existing stack

Client Overview

Albatross is the foundational AI platform for real-time discovery, pioneering perception models that adapt to behavior in real-time. Today, Albatross processes billions of user interactions every month and runs multiple model training cycles daily for hundreds of millions of items. The team operates at the intersection of applied ML research and production-grade inference, demanding rapid model iteration, reliable serving, and disciplined compute economics.

The Challenge

Running on reserved hyperscaler GPUs, Albatross hit three problems at once: low GPU utilisation between bursty retrains meant paying for idle capacity, access to higher-end GPUs was inconsistent, and long-running R&D experiments didn't fit the same pricing model as production jobs. Their ML pipeline also had integration requirements that generic cloud APIs couldn't meet off the shelf.

The Solution

Lyceum's engineering team built a custom integration that drops Lyceum GPUs directly into Albatross's existing Kubernetes cluster — no new cloud stack, no software rewrites:

Serverless GPU for production retraining. Albatross deploys containerised training jobs that spin up on-demand, execute, and release the job immediately once done. They only pay for active GPU-seconds, eliminating idle-time waste across their training cycles.

Custom integration built by Lyceum. Lyceum's engineers built custom software that provisions AWS EKS Hybrid Nodes on-demand, plugging Lyceum GPUs directly into Albatross's existing EKS cluster. Spinning up a hybrid node is a one-click action — no manual installation or configuration of the specialised networking software that would normally be required. The result: Albatross replaced AWS GPU instances with Lyceum ones at lower cost, with zero changes to its cloud setup or software stack.

The Results

Better GPU economics without changing the stack. Albatross kept its existing tooling, moved to pay-per-use GPU economics, and gained reliable on-demand access to higher-end GPUs.

Lyceum lets us match infrastructure to each workload instead of overpaying for one rigid setup. The result is better economics, faster iteration, and a genuinely collaborative way of working with a team we can evolve alongside.

— Dr. Kevin Kahn, Co-Founder and CEO, Albatross

A tailored integration, shipped in days. Rather than leaving Albatross to adapt to a generic API, Lyceum's engineers built the integration straight into Albatross's ML pipeline. It's live today, with production deployment on track.

What stood out was how fast Lyceum turned our requirements into a working integration that fit our ML workflow. No quota games or capacity planning, just a direct path to production.

— Dr. Matteo Ruffini, Co-founder and Chief Scientist, Albatross

Engineering time back on models. With capacity and integration handled, Albatross's team ships model improvements rather than chasing GPUs.

With Lyceum, we focus on model development and iteration instead of worrying about GPU availability or infrastructure pitfalls.

— Johan Boissard, Co-founder and CTO, Albatross