Catalog
concept#AI#Machine Learning#Architecture#Data

Foundation Models

General concept of large pretrained AI models that serve as a base for various applications.

Foundation models are large pretrained AI models that serve as a general-purpose base for many downstream tasks.
Emerging
High

Classification

  • High
  • Technical
  • Architectural
  • Intermediate

Technical context

API gateways and inference endpointsData platforms and feature storesMonitoring and observability tools

Principles & goals

Treat models as reusable base layers.Prioritize data quality and diversity.Plan governance, security and monitoring from the start.
Build
Enterprise, Domain

Use cases & scenarios

Compromises

  • Misuse via generation of misleading content.
  • Privacy or licensing violations from training data.
  • Excessive dependency on third parties or models.
  • Use version control for models and training artifacts.
  • Establish continuous monitoring and drift detection.
  • Define clear responsibilities for governance.

I/O & resources

  • Pretrained model weights
  • Curated domain dataset
  • Infrastructure for training and inference
  • Fine-tuned models for product features
  • Evaluation reports and test sets
  • Operational artifacts (pipelines, monitoring)

Description

Foundation models are large pretrained AI models that serve as a general-purpose base for many downstream tasks. They are trained on broad data collections and adapted via fine-tuning or prompting for specific applications. Their adoption requires careful governance, data strategy, and security considerations.

✔

  • Faster product development via pretrained capabilities.
  • Improved generalization across tasks.
  • Efficiency gains through transfer learning and reuse.

✖

  • High compute and memory requirements for training and inference.
  • Dependence on large, often proprietary datasets.
  • Potential biases and undesired behaviors.

  • Inference latency

    Average response time for model requests, important for UX.

  • Requests cost (Cost per request)

    Operational cost per request including infrastructure and model access.

  • Accuracy / domain-specific metrics

    Performance metrics tailored to specific tasks (e.g. F1, BLEU).

Chat assistant with the GPT family

Use of large generative foundation models to answer user questions in real time.

Document analysis with BERT-based models

Fine-tuned models for classification and extraction from business documents.

Code generators based on large models

Automatic code suggestions and templates from pretrained models adapted to developer workflows.

1

Define use case and specify data requirements.

2

Select a pretrained model and set evaluation criteria.

3

Develop fine-tuning or prompting prototype.

4

Run rollout with monitoring, tests and governance.

⚠️ Technical debt & bottlenecks

  • Monolithic model deployments without modularization.
  • Missing automation for retraining and rollbacks.
  • Insufficient documentation of training data and pipelines.
Compute resourcesData acquisitionLatency for interactive applications
  • Generating legally problematic content without moderation.
  • Using sensitive internal data for unchecked fine-tuning.
  • Use in safety-critical contexts without robustness tests.
  • Overestimating generalization capability on niche domains.
  • Underestimating ongoing operational costs.
  • Poor measurement of quality metrics in product context.
Machine learning and model architecturesData engineering and feature engineeringDevOps for ML infrastructure (MLOps)
Scalable inference infrastructureData quality and managementSecurity and governance requirements
  • • Budget limitations for training and inference
  • • Legal and licensing constraints for training data
  • • Operational overhead for monitoring and security