LLM Features That Ship: Extraction, Generation, and Classification (alex-jacobs.com)

After years of shipping LLM-powered features, the author presents a pragmatic framework: the reliably deployable LLM use cases fall into three categories—structured data extraction, controlled content generation/summarization, and categorization/classification. Extraction covers PDF→JSON, noisy OCR/table extraction, API response mapping, and pulling actionable items from feedback streams—tasks where LLMs handle ambiguity and format variance far better than brittle regex pipelines. Generation is valuable when bounded—report drafts, meeting summaries, and docs—requiring explicit schemas, length/tone constraints, and validation layers to avoid unconstrained hallucination. Classification lets teams create declarative classifiers without labeled datasets, enabling fast iteration for business categories that would otherwise need heavy ML pipelines. The post also sketches production realities and technical guardrails: prefer buy vs build for high-volume extraction (e.g., Reducto), enforce structured output with tools like Pydantic AI or Outlines, treat prompt engineering as an optimizable component (DSPy), and invest in evals/observability. Anti-patterns include replacing deep domain experts, sub-100ms real-time use, and tasks demanding perfect accuracy. Practically, start with classification—clear metrics, forgiving failure modes, and easy human-in-the-loop fallbacks—and design features to augment workflows rather than replace them, focusing on where LLMs remove tedious, error-prone human work.