The Premise
Most enterprise LLM deployments are decorative. A chat interface bolted onto a document store — ask a question, copy the answer, repeat. Useful, but not automation.
What I’ve been building is different: LLMs as operational components that ingest structured inputs, apply business logic, and return results that flow directly into pipelines. No human in the loop for routine decisions. No chatbot facade.
The Core Insight
Language models perform significantly better when given an explicit model of the organisational context they’re operating in.
Every organisation has a vocabulary, a hierarchy, and a set of unstated rules about what “correct” looks like. Generic prompts produce generic output. Prompts that encode institutional knowledge — terminology, decision criteria, edge cases — produce something closer to a junior specialist.
I’ve been developing context-aware prompt architectures that treat institutional knowledge as a first-class input, injected before any task-specific content.
What’s Been Built
Document Ingestion Agents
Custom LLM wrappers for processing unstructured enterprise documentation — internal policy files, shift rotation rules, operational reports. The agent reads, identifies logical structures, and outputs normalised JSON for downstream pipelines.
Enterprise documents are written for humans, not parsers. They use abbreviations the model hasn’t seen, reference tables stored elsewhere, and assume context that isn’t in the text. The solution: context injection layers that pre-load the relevant organisational vocabulary before each document is processed.
Structured Reasoning Chains
Multi-step pipelines where the LLM acts as a logic engine: given a set of constraints (labour law rules, operational parameters), evaluate cases and return a structured decision with reasoning.
The critical architectural decision: separate reasoning from formatting. A prompt that asks “analyse this AND return JSON” degrades both. Two prompts — one for reasoning, one for extraction — produce significantly better results at both tasks.
Status
Active research and personal exploration. The WFM Lakehouse and the studio work were both shaped by what I’ve learned here. The longer-term direction is autonomous pre-processing agents that slot into existing data pipelines without requiring architectural rewrites.
Stack: gpt-4o for reasoning-heavy tasks · gpt-4o-mini for high-volume extraction · LangChain for orchestration (though the raw API wins when abstractions add more complexity than they remove) · Python throughout.