I built Galuchat, a compact reverse-geocoding system that works without an external geocoding server or API.

The same basic approach can be used in:

Browsers WebMCP ESP32 RP2040 / RP2350 JavaScript / C++ / Java / Python

The basic idea is to store the administrative-boundary data locally and perform the lookup directly on the device, instead of sending every coordinate to an external API.

Rather than storing the original polygon data as-is, Galuchat converts administrative regions into a very large raster segmentation image.

The raster uses resolutions ranging from roughly 10 m to 1 km per pixel. It is then compressed into a binary representation small enough to be embedded in MCU flash or downloaded as a single dataset by a browser.

Some actual dataset sizes are:

Japan-wide administrative boundaries ~1.1 km resolution: ~52 KiB ~111 m resolution: ~479 KiB ~11 m resolution: ~4.47 MiB Place-name dictionary: ~31 KiB Worldwide administrative boundaries ~1.1 km resolution: ~2.81 MiB ~111 m resolution: ~20.45 MiB Place-name dictionary: ~561 KiB Japan-wide administrative + town-block / small-area data ~11 m resolution: ~24.07 MiB Place-name dictionary: ~1.66 MiB Total: about 26 MiB

This makes several kinds of use possible:

Embedding a Japan-wide administrative-area dataset of only tens or hundreds of KiB directly into a small MCU Running more detailed offline reverse geocoding on ESP32-class devices with larger flash Downloading a worldwide dataset once in a browser and then processing large numbers of coordinates locally without external API calls Using the detailed Japan dataset to resolve locations below the municipality level entirely inside the browser

For example, a GPS-enabled embedded device can determine the administrative area for its current latitude and longitude without Wi-Fi or cellular connectivity.

The same applies in a browser. Once the GIS dataset has been loaded, individual coordinates do not need to be sent to an external service. Large GPS logs can therefore be processed locally without per-request API calls or service rate limits.

This means that even worldwide administrative-boundary data can be embedded in MCU flash or downloaded into a browser and queried locally.

WebMCP can use the same local dataset as well, so an AI agent can call the reverse geocoder directly inside the browser without relying on an external geocoding service.

Demo and documentation: https://nyatla.github.io/galuchat/

Source: https://github.com/nyatla/galuchat-gis-sdk

I originally started this project as an experiment to see how small a practical geographic lookup dataset could be made.

It eventually became small enough to run on microcontrollers, so I extended the same approach to browsers and WebMCP as well.

Read your own writes, off the primary

Anthropic 3Q26 Profit over $1B: The Anthropic IPO Financials Sneak Peak

Show HN: Open-source transaction enrichment dataset

Show HN: Serverless reverse geocoding for WebMCP, ESP32, and RP2040

Proxmox Decisions That Are Painful to Change Later

Chromium, Chrome, Edge, Opera, Vivaldi: Multiple Vulnerabilities

Detecting a pebble at 60M km at 99% of light speed

ChatGPT Linux GUI version updates redownload all files, require 1.4GB download

Xfinlab – a financial intelligence API with an MCP server for AI agents

BGP hijack infecting networks caused by a comedy of errors

Academia as a d-index measuring contest

Dog Age Calculator

9000 RPM is faster than your screen refreshes

Using jq to format JSON on the clipboard

Show HN: Local Personal Financial Assistant

Chladni Plate

An update on our dispute with Disney

Autopsy of an N=1 Cybernetic Therapy: Multimodal AI for Grief (Case Report)

Mistral trains on user input by default, except on enterprise tier

Show HN: Review Markdown files like a document

B200 Attention Kernel from Scratch to Near-SOTA in 60 Diagrams

Ask HN: Where do you get your AI (technology) news?

Csveee – parsing CSV at up to 192 GB/s in Rust

A call for collective action on cyber defense

Uber is laying off 10% of staff

Adversarial Collaboration in Philosophy

Why I prefer using LLM API aggregators over subscription services

Check if a file was made with Claude

What to Do While Waiting for AI Code Assistants

Building a personal book library with NotebookLM and Claude Code

Noviz AI Agent for ERPNext – A Reasoning Layer for ERPNext