Hi! I'm working on my side-project Meshfox and created a simple static site generator for mindmap-style docs based on it. You can use meshfox editor with one of two example templates from repo or create your own one. You can check result on project own site. Project is still in beta phase, but very close to v1.0 checkpoint.
Scenic pedestrian routing via LLM + custom Valhalla costing
Built an MCP server that takes natural-language walk requests and returns a GPX/GeoJSON route. All data from OpenStreetMap, self-hosted via Docker.
The approach:
OSM features are scored by semantic similarity to the query using pgvector + local embeddings (Ollama / nomic-embed-text)
A 50 m/cell heatmap is built over the route bbox; each feature contributes a quadratic kernel weighted by sim⁴
The heatmap is passed to a patched Valhalla instance via a custom `scenic_pedestrian` costing mode — edges in hot zones cost up to 10× less
The route is scored by sampling heat every 50 m. If below threshold, peak waypoints are extracted via PCA + flood-fill + bitmask DP and the route is re-planned through them
Stack: Go, Valhalla (patched), PostGIS + pgvector, Ollama, Docker Compose. Nightly OSM updates from Geofabrik.
Repository: https://github.com/orofarne/scenic-routing-mcp
Algorithm writeup (the interesting part — kernel design, aggregation choices, Valhalla protobuf packing, Gini-based waypoint decision): https://github.com/orofarne/scenic-routing-mcp/blob/main/docs/algorithm.md