AI Geospatial Platform
Vyoman Earth Intelligence
Turns Luxembourg's public geospatial data and drone imagery into an evidence-backed, human-reviewed site briefing — in minutes, not a GIS analyst's week.
The problem
Site analysis for architectural and planning projects in Luxembourg means manually querying three separate government WMS servers, stitching together orthophoto tiles, interpreting LiDAR elevation data, and cross-referencing drone imagery — a GIS analyst's week of work per site. There is no tool that assembles this into a human-readable briefing automatically, and no one is building one for the Luxembourg government data stack.
How it works
- Live data ingestion from three public sources: Geoportail WMS orthophotos (10cm resolution aerial imagery), LiDAR Digital Elevation Model (terrain height data), and the operator's own Vyoman drone imagery for sub-metre site-level detail not available in government data.
- Evidence-packet architecture: the same pattern used in ComplyLens, Disha Analytics, and Soul Spark — raw data is transformed into structured, citable evidence before any AI call touches it. The AI brief cites the packet; the human reviewer inspects the packet directly.
- Seven workflow tabs cover the full site intelligence pipeline: AOI (area of interest) drawing, orthophoto fetch, DEM terrain analysis, drone imagery overlay, AI briefing generation, human review with annotation tools, and report export.
- PyDeck 3D terrain renderer: renders LiDAR DEM data as an interactive 3D surface alongside the 2D map view — giving planners a spatial intuition for elevation changes that flat maps lose.
- Human review workflow: draft → reviewed → approved lifecycle with DuckDB audit trail. Human annotators can add site-specific observations before the briefing is finalised. The AI draft is never the final output.
- Local-first with Ollama — all LLM inference runs on-device. A deterministic mock fallback ensures the pipeline works even without a running Ollama instance, making it testable without GPU access.
- ReportLab PDF export: the final approved briefing exports as a structured PDF with embedded map snapshots, data source citations, and reviewer annotations.
The journey
A local prototype that turns three government data sources and the operator's own drone imagery into a structured, human-reviewed site briefing — using the same evidence-packet architecture as the other products in this portfolio.
Define the site — draw an AOI, the pipeline fetches everything
The analyst draws an Area of Interest polygon on a Leaflet map. The pipeline immediately fetches Geoportail WMS orthophotos at 10cm source resolution for that tile, pulls LiDAR Digital Elevation Model data for the same bounds, and lists any Vyoman drone imagery that covers the AOI. Three data sources, one interaction.
Terrain analysis — 2D map and 3D surface, side by side
LiDAR DEM data is rendered as an interactive 3D terrain surface via PyDeck alongside the 2D orthophoto map. Elevation profiles, slope analysis, and shadow modelling give planners spatial intuition that flat maps cannot — especially relevant for Luxembourg's mixed urban and hillside topography.
Evidence packet assembled — AI generates the brief from structured data
All fetched data is transformed into a structured evidence packet: georeferenced extents, elevation statistics, data source metadata, and image descriptors. Ollama (local LLM) reads the packet and generates a site briefing that cites specific evidence entries. The human reviewer can inspect the raw packet — not just the AI summary.
Human review, annotation, and approved export
The AI draft enters a review workflow: draft → reviewed → approved. Reviewers add site-specific annotations, flag uncertainties, and confirm data quality. Every state change is logged in DuckDB with a timestamp and reviewer ID. Only approved briefings export. The final PDF includes embedded map snapshots, data source citations, and all reviewer annotations.
Highlights
- Live-fetches Geoportail WMS orthophotos at 10cm source resolution
- LiDAR DEM terrain rendered in 3D via PyDeck alongside 2D map
- Evidence-packet architecture — AI cites structured data, not raw text
- Human review workflow: draft → reviewed → approved + DuckDB audit trail
- Local-first: Ollama inference + deterministic mock fallback, zero cloud API keys
- 7-tab pipeline: AOI → fetch → terrain → overlay → brief → review → export
Stack
Signature features
Evidence-packet architecture
Raw WMS, LiDAR, and drone data are structured into a citable evidence packet before any LLM call — the same pattern used in ComplyLens, Disha, and Soul Spark.
Live Geoportail WMS ingestion
Fetches Luxembourg government orthophotos at 10cm source resolution on demand — no pre-downloaded tiles, no stale imagery.
2D map + 3D terrain
PyDeck renders LiDAR DEM data as an interactive 3D surface alongside the 2D map view, giving planners spatial intuition that flat maps lose.
Human review workflow
Draft → reviewed → approved lifecycle with DuckDB audit trail. Human annotators can add site observations before finalising — the AI draft is never the final output.
Local-first Ollama + mock fallback
All LLM inference runs on-device. A deterministic mock fallback makes the full pipeline testable without a running Ollama instance.
Vyoman integration
Every Vyoman aerial shoot generates imagery that feeds directly into this intelligence pipeline — closing the loop between the drone operation and the site analysis tool.
By the numbers
3
live data sources
7
workflow tabs
10cm
ortho source resolution
0
cloud API keys
Why it matters
Vyoman Earth Intelligence demonstrates the same evidence-packet architecture pattern used across Disha, ComplyLens, and Soul Spark — applied to geospatial data. The project proves that sophisticated GIS-style analysis is accessible as a local prototype with no cloud dependencies. Built as an extension of the Vyoman Aerials drone operation: every flight generates imagery that feeds directly into this intelligence pipeline. The natural next step is integrating Vantage mission data so that flight plans and site briefings are generated from the same canonical data layer.