Foundation models and language systems

Large pre-trained models have become platforms: general-purpose components that are adapted with small amounts of task data, wrapped with retrieval and tools, and composed into pipelines. That shift moves the hard questions from “can a model do X” to “can a system built on models do X reliably, at a known cost, with failure modes we understand”.

We study foundation models from the systems side. Our interest is less in training the next largest model and more in the engineering around them: adaptation strategies, retrieval and grounding, evaluation harnesses that survive model upgrades, and the measurement discipline that tells a real improvement from noise.

What we work on

• Adaptation choices. Comparing fine-tuning, retrieval augmentation, prompting, and tool use on the same task with the same held-out data, so the choice is driven by evidence.
• Grounding and provenance. Keeping an auditable trail from a model’s answer back to the sources and the code version that produced it.
• Evaluation that lasts. Regression suites and point-in-time test sets that catch silent changes in hosted models.