Home · Features · AI / ML

AI that never phones home.

Gemma 3 / Phi-4 through a llama.cpp sidecar, SAM2 for segmentation, YOLO-E for FMV and still imagery. Natural-language queries, masks, and georeferenced detections — enclaves without a cloud.

Gemma 3 / Phi-4 llama.cpp sidecar NL→SQL (DuckDB) SAM2 segmentation YOLO-E detection 100% local inference

Local-first architecture

A Python sidecar, spawned on demand

C#/WPF host → FastAPI orchestrator → llama.cpp with Gemma GGUF. Lazy-start: sidecar only when you use AI. Every endpoint 503s cleanly if off.

Geospatial chatbot

Gemma 3 / Phi-4 via llama.cpp — grounded in local data, not a hosted API.

NL→SQL for DuckDB

Drafts SQL with schema context (incl. ST_DWithin) — you review in Query Builder before Run.

SAM2 segmentation

Mark an area on a COG → GeoJSON polygon overlays.

YOLO-E detection

Imagery FeatureSets + FMV frame detection → CoT / TAK broadcast.

Human-in-the-loop

AI suggests. The analyst approves.

NL→SQL fills the editor — never executes blind
Class labels + confidence set before detection
FMV AI detection off by default
Weights in %APPDATA%/3DMapExplorer/models/

AI orchestrator · localhost onlyAiOrchestratorHost → sidecar :8004
  ├─ /v1/chat    → llama.cpp :8080
  ├─ /nl2sql     → DuckDB SQL
  ├─ /segment    → SAM2 on COGs
  └─ /detect/cog → YOLO-E FeatureSet

GET  /api/ai/capabilities
POST /api/ai/chat · /nl2sql · /segment
● 100% local · no cloud AI calls

Bring the models to the data. Not the reverse.

Drop the weights on disk, flip on the sidecar — chat, NL→SQL, segmentation, and detection inside the COP.

Get Started See imagery intelligence