project · 2024-2025

Natural-language location planner

Designed a planner-executor system that converts natural-language location requests into grounded, multi-step tool calls and explainable results.

#agentic #llm #rag #azure-openai #full-stack #fastapi #nextjs

The problem

Location questions often combine several intentions—finding a place, meeting a constraint, and planning a route. A reliable experience needs to separate language understanding from factual, tool-based decisions.

Approach

Used an LLM planner to produce a typed, inspectable plan.
Used deterministic execution for search, routing, filtering, and ranking operations.
Returned source-aware results so users can understand why a recommendation was made.

How it works

The planner turns a multi-intent request into a typed plan. A deterministic executor then runs the individual tools and assembles the result. This replay uses public example locations and synthetic results.

planner + executor demo · multi-intent query

press play to watch the query parse, plan, and execute.

mapidle

Demo basemap: open-source CartoDB / OpenStreetMap tiles, no API key required. The demo uses public tiles so the portfolio remains self-contained and key-free.

Outcome

The planner-executor pattern provided a useful balance between natural interaction and reliable, grounded behaviour.

← all projects