Foundation models
Pretraining and post-training open-weight large models for specialized domains and languages.
- Continued pretraining
- Post-training
- Alignment
- Domain adaptation
Multimodal understanding
Models that take in text, images, audio and documents together, because information at work rarely comes in a single format.
- Vision-language
- Speech
- Document understanding
- Cross-modal retrieval
Lightweight models
A large model's results from a smaller, faster and cheaper one, on a server or on the device itself.
- Distillation
- Quantization
- Sparse models
- Edge inference
Working with sensitive data
Models that run fully on-premise and learn from sensitive data without it ever leaving its source.
- On-premise LLMs
- Federated learning
- Differential privacy
- Air-gapped deployment
Learning from limited data
Methods for few labels, noisy records and conditions that keep changing, which is how data arrives in most projects.
- Transfer learning
- Self-supervision
- Synthetic data
- Continual learning
Controlled agents
Planning, permissions and a check at every step for agents that take actions in live systems.
- Planning
- Step verification
- Permission models
- Agent evaluation
Interpretability & evaluation
Measuring what a model actually does and understanding why. Auditing, debugging and deciding whether to trust a model all depend on it.
- Attribution
- Mechanistic interpretability
- Benchmark design
- Failure analysis
Experiments & simulation
Models that combine domain knowledge with what they learn from data, standing in for slow simulations and costly experiments.
- Surrogate models
- Scientific ML
- Digital twins
