An AI-assisted workflow that qualifies B2B travel companies from industry relevance, business model and API or infrastructure orientation.
The problem
The first version delegated both classification and reasoning to the model, producing variable results across runs.
The solution
I moved stable segmentation and next-action rules into code. The model remains responsible for fit score and concise qualitative reasoning.
What I built
I built a Node.js workflow that classifies company signals, uses JSON Schema for output and sends OpenAI only the context it needs.
How it works
Company data ? business rules ? segment and signals ? LLM ? fit score and qualitative reasoning.
Engineering choices
Langfuse and OpenTelemetry provide tracing, prompt versioning, datasets, experiments and automated evaluators. An explicit rubric makes prompts and results comparable.
Outcome
In the V3 benchmark the workflow achieved 15 correct classifications out of 15. The main outcome is separating deterministic business rules from probabilistic reasoning.
Gallery
