Three Patterns of Successful Project Implementations with LLM Under the Hood

Rinat Abdullin25 May 2024

Problem

A company or an AI consultancy wants to bring LLMs into existing business processes at a medium or large enterprise. Without a known project shape it is hard to estimate the work, to split it between the LLM team and the integration team, or to refuse a risky project before months are lost.

Example from the article: A supplier of electronic components fell under sanctions, and competitors wanted its customers. The competitors had to collect all datasheets, extract structured specifications from them in bulk and match them with their own products.

Solution in brief

The speaker groups LLM projects into three patterns: data extraction at scale, document assistants (AI search), and AI platforms. Each pattern needs a different amount of client communication and team involvement. On top of the patterns he gives a delivery framework: map client problems to cases you already know, uncover risks early, and rank requests by seven criteria.

Questions this case answers
  • How to choose the right architecture for an LLM project that extracts structured fields in bulk, answers questions over a document base, or ties an LLM into business processes
  • How to estimate the work and split it between a model team and an integration team before months are spent on a risky LLM project
  • How to choose an architecture pattern for an LLM project: data extraction, document assistant or AI platform
  • How to scope an LLM data-extraction project so that it is easy to agree on and to hand over for integration

Tools and methods

  • Data extraction at scale — documents in, checkable structured data out, as a stable interface between teams
  • Knowledge mapping and a classifier agent — route a question to an agent restricted to the relevant files
  • POC with 3-5 feedback iterations — run on a dump of client data
  • Case matrix and seven prioritization criteria — pick which client requests to take
  • Camelot, Unstructured, Tesseract, GPT-4 Vision — PDF text and table extraction
Continue reading — it's free

Google or a one-time email code. No password.