AskTheMap

Inspiration

We were inspired by a simple problem: people already coordinate trips, meetups, and “where should we go?” decisions inside chat apps, but map tools usually live in separate apps and tabs. We wanted to bring map intelligence directly into the conversation, so users could ask for routes, nearby places, and travel insights in natural language without leaving WhatsApp.

What it does

AskTheMap lets people ask map questions directly in WhatsApp and get instant, AI-powered replies. It understands natural-language requests, calls mapping tools when needed (Google Maps/Mapbox), and returns useful results like:

Directions and route options
Nearby places and points of interest
Travel-time and location insights
Group-friendly planning help inside chat
It runs as a backend gateway, so conversations stay in WhatsApp while the system handles LLM reasoning, tool calls, and reliable message processing behind the scenes.

How we built it

We built AskTheMap as a long-running Express gateway connected to WhatsApp Web. Incoming messages are processed through an LLM orchestration layer, and map-specific requests are executed through an internal MCP-style tool registry.

Core architecture:

WhatsApp Web client for real-time chat I/O
Express service with health endpoint for observability
Queue-based background processing for stable, ordered handling
Message dedupe + debounce to reduce repeated/fragmented inputs
Gemini-powered agent with MCP tool calls
Google Maps and Mapbox integrations for geocoding, routing, and isochrone-style capabilities

Challenges we ran into

Balancing speed vs. reliability when tool calls timeout or fail intermittently
Preventing duplicate replies from repeated webhook/chat events
Handling rapid multi-message bursts without losing user intent
Keeping tool usage bounded so the model doesn’t loop indefinitely
Managing API-key compatibility and environment differences across team setups

Accomplishments that we're proud of

Built a working WhatsApp-first AI mapping assistant end to end
Integrated LLM reasoning with real map tooling (Google Maps + Mapbox)
Added robust message handling with queueing, deduplication, and debounce
Implemented MCP tool-call guardrails (round limits, retries, fail-safe final answers)
Designed chat controls (/bot on, /bot off, /reset) for safer real-world use
Shipped operational visibility via /healthz and structured logging
Kept setup beginner-friendly with clear env templates and quick-start docs

What we learned

Reliability features (queueing, dedupe, debounce) are essential for chat UX, not optional.
LLM quality alone is not enough; tool orchestration and guardrails determine real usefulness.
Bounded MCP loops, retries, and timeout handling prevent runaway behavior in production.
Keeping map tools server-side improves safety, control, and observability.
Natural-language chat can be a practical interface for route planning and place discovery when context is preserved.

What's next for AskTheMap

Per-person memory in group chats. Add individual memory so AskTheMap can remember each person’s preferences and routines (food tastes, budget, neighborhoods, typical availability, mobility constraints), while keeping those preferences separated per user in the same group chat.
Turn it into a general API (not just a WhatsApp bot account). Evolve AskTheMap into a reusable service with a clean API—so it can power multiple clients (WhatsApp, web, mobile, other chat apps) instead of being tied to a single WhatsApp “secondary account” setup.
More map-understanding skills. Expand the skill set beyond basic search/directions into richer “map intelligence,” like better place comparison, area recommendations, isochrone-based suggestions, multi-stop planning, and smarter context-aware local recommendations.