Monitor mining from space. Confirm it on the ground.
Lalibela helps government agencies find suspected illegal mining in satellite imagery and send an inspector to check it. Every finding stays a suspicion until someone has verified it on site.
31°E to 49°E · 3°N to 15°N
Oromia: Shakiso, Adola, Kenticha
Dima
Asosa
Adi Dairo
Sample survey areas
AI proposes. People decide.
Illegal and unregistered mining is hard to see from an office. Lalibela gives a regulator's team one place to look at imagery, review what the models flag, send someone to check, and keep a record of what was found.
Satellite imagery comes in
An operator selects a survey area and loads imagery with its capture date.
Models analyze it
Qwen, Gemini and Claude can be run side by side on the same image. Their findings appear as highlighted areas.
An operator reviews each finding
Nothing becomes a case until a person has looked at it and marked it reviewed.
A case goes to a field inspector
The inspector sees the assignment, navigates to the site and must pass a GPS check before starting.
The inspection is recorded and reviewed
Outcome, observations, photographs and location are stored with the case. An operator reviews the submission and closes the case.
Reports go to decision makers
Filter by region, date, inspector or model and export to PDF or Excel.
Follow one case from imagery to closure
This is a working model of the Lalibela case flow using sample imagery near Shakiso. Run the analysis, review the findings, assign an inspector, then check in on site. The model results are simulated and nothing leaves your browser.
Analyze imagery
Sample capture of Shakiso · Survey A. Choose the models to compare, then run the analysis.
This demo returns fixed sample findings. No image is sent to Qwen, Gemini or Claude.
Four roles, one shared record
Each person sees the screens they need. The prototype shows how access is divided. A live deployment enforces it on the server.
Administrator
Creates and edits users, assigns roles and switches them on or off. Chooses which AI models are available and sets the GPS verification radius.
System Operator
Previews imagery, runs the analysis, reviews findings, creates cases, assigns inspectors and closes cases after reviewing the inspection.
Field Inspector
Sees every case assigned to them. Opens case details, navigates to the site, verifies location, adds photographs and observations, and submits the inspection from a mobile view.
Dashboard Viewer
Read-only access to the overview, map, case register, case history and reports, for managers and partner agencies.
A case moves one way
Cases go from New to Assigned to Under Inspection to Inspection Submitted to Closed. There are no hidden steps backward. What the inspector found is stored separately, as one of three outcomes: confirmed violation, no violation found, or inconclusive with follow-up required.
| Step | Who | What is checked |
|---|---|---|
| Create a case | Operator | A reviewed finding, evidence reference and observations |
| Assign | Operator | An active inspector and an inspection target |
| Start inspection | Assigned inspector | A fresh on-site GPS check |
| Submit inspection | Assigned inspector | Outcome, observations, evidence and a fresh location |
| Close | Operator | Submission reviewed and a closing note recorded |
- The inspector must be thereA location counts only if the fix is under two minutes old and the distance plus the reported accuracy fits inside the radius. Try it in the walkthrough.
- Models disagree in the openEach model is run on its own. Differences and failures are shown, and a score from a model is never treated as the probability of a violation.
- GPS is a safeguard, not proofA passed check allows an action. It does not prove presence, so evidence review and audit history stay part of the process.
Reports for people who were not in the field
Filter by location, date range, inspector or AI model. The summary and the export follow the filters, so what a director sees on screen matches the file they send on.
| Export | Contents |
|---|---|
| PDF summary | Filtered totals: suspected findings, cases created, areas inspected, cases by status and inspection outcomes |
| Excel workbook | A summary sheet, a case sheet with inspector, status, outcome, GPS coordinates and GPS source, and a sheet of findings still waiting for a case |
What exists today, and what a deployment adds
Lalibela is currently an interactive prototype built to show the experience and the rules. It has no shared database, no live model connection and no real investigations in it. A design for taking it to production is written and ready to be planned.
| Capability | In the prototype today | For a live deployment |
|---|---|---|
| Sign-in and roles | Sample accounts and interface-level permissions | Managed identity, multi-factor sign-in and server-side authorization |
| Cases and history | Held in the browser for one session | Shared database with an audit trail |
| Imagery upload | JPEG, PNG or WebP previews | GeoTIFF ingestion and private storage |
| AI analysis | Simulated Qwen, Gemini and Claude results | Model adapters running real inference, with human review kept |
| Field app | Responsive mobile web view | Native Android and iOS app with offline drafts |
| Location checks | Browser GPS or clearly marked simulation | Server-side verification and anti-spoofing signals |
| Reports | PDF and Excel created in the browser | Authorized, reproducible server exports |
The stack above is the recommended design, not something already built. Model choice should follow a local evaluation on licensed imagery.