Language & Generative AI
This is the layer most businesses actually touch — large language models and the patterns for pointing them at your own work.
LLMs and Foundation Models
A foundation model is trained once, at great expense, on a very broad corpus, then adapted to countless downstream tasks. Large language models are the text-and-code version. Practically, this means you rent capability rather than build it: the base intelligence is a commodity, and your advantage comes from the data, workflows, and judgment you wrap around it.
Fine-Tuning
Fine-tuning continues training a base model on your own examples so it internalizes a format, tone, or classification scheme. It's excellent for consistency and for compressing long instructions into the model itself. It is not a way to teach the model facts that change — for that you want retrieval. Fine-tuning costs money, requires curated examples, and has to be redone as the base model evolves.
RAG — Retrieval-Augmented Generation
RAG is the default pattern for making a model knowledgeable about your business. You index your documents, retrieve the passages relevant to each question, and hand them to the model along with the question. The model answers from supplied evidence rather than memory — which dramatically reduces hallucination, lets you cite sources, and means updating knowledge is just updating documents.
- Keeps answers grounded in your actual policies, contracts, and records
- Content updates instantly — no retraining
- Enables citations, which is what makes answers auditable
Agents and Tool Use
An agent is a model given tools — search, a database query, an API call, a calculator — and the latitude to decide which to use and in what order to reach a goal. This is where AI stops answering and starts doing. It's also where risk concentrates: every tool an agent can call is an action it can take wrongly, so scope permissions narrowly and log everything.
NLP, Computer Vision, and Multimodal
Natural language processing covers understanding and generating text; computer vision covers images and video. Multimodal models handle both at once — reading a screenshot, describing a photo, extracting a table from a scanned invoice. For most operational work, multimodal capability is what unlocks the messy real-world documents businesses actually run on.
Start with RAG over your own documents. Add fine-tuning only for consistency problems, and add agents only where you can constrain and audit the actions.