Cutting manual location entry from 50% to 2% with an LLM pipeline

Location input looks simple until users start entering things like:

“Dubai airport T3”
“Marriott near DXB”
“Airport terminal 3 Dubai”

For a global booking platform, these messy, incomplete, and multilingual inputs created a major operational problem.

At one point, ~50% of location entries required manual search or verification.

The Problem

A traditional flow looked like:

User Input → Map Search → Results → Location

The problem wasn't just finding coordinates.

The system first needed to understand what the user meant.

The Solution

I introduced an LLM-powered normalization layer using:

OpenAI + Mapbox + LangSmith

The architecture became:

Raw Location ↓ OpenAI ↓ Structured Location ↓ Mapbox ↓ Resolved Location ↓ Booking Workflow

The responsibilities were deliberately separated:

• OpenAI → understand and normalize human input
• Mapbox → resolve the geographic location
• LangSmith → trace and observe the LLM workflow

The LLM wasn't responsible for generating coordinates or making every decision.

It acted as an intelligence layer between messy human input and deterministic mapping infrastructure.

The Result

Before

~50% → Manual location search

After

~1–2% → Manual location search

That's roughly a 96–98% reduction in manual location intervention.

The Key Lesson

The interesting part wasn't simply adding an LLM.

It was deciding where the LLM should be used.

Let probabilistic systems understand.
Let deterministic systems resolve.

This pattern works well whenever humans provide messy information but downstream systems need structured, predictable data.