Aquarium
Aquarium is the AI model management platform used internally at AISG, covering dataset management, training experiment tracking, model version control, deployment monitoring and other ML lifecycle stages. It is not a standalone product but rather AISG's "internal MLOps system".
📖 What it is
Aquarium's functional modules:
- Dataset management: versioning, annotation, distribution analysis
- Experiment tracking: training metrics, hyperparameters, checkpoints
- Model registry: with version rollback support
- Deployment monitoring: performance and drift monitoring for in-production models
The design resembles a combination of MLflow + Weights & Biases + DVC, but customised for AISG's own workflow.
🤖 Relation to AI
Aquarium's role inside AISG: giving AIAP apprentices, the SEA-LION team, and every AI project a shared ML engineering foundation.
The value lies in:
- Apprentices don't have to set up experiment tracking from scratch each project
- Checkpoint and dataset management for major projects like SEA-LION follow a unified standard
- Datasets and components can be reused across projects
🇸🇬 Relation to Singapore
Aquarium reflects AISG's internal "engineering rigour" — a national-level AI institution needs engineering infrastructure, or labour costs get eaten up by infrastructure-building.
In the seven-lever framework:
- Lever 1 (infrastructure): the foundation of AISG's internal ML engineering capability
A take: Aquarium is not AISG's flagship external product, but it is the engineering foundation that lets AISG keep delivering at high tempo (SEA-LION, TagUI, PeekingDuck, and more).