wispayr / vision

Alice Vision Unit — v1 YOLO curation, MCP-native, drop-in on any Linux host.

Train per-site vision models without a dataset-ops team. A vision unit pulls snapshots from your cameras, auto-labels with a YOLO11x teacher, deduplicates with perceptual hashing, fine-tunes an 11n student, and exports ONNX FP16 for the edge — all driven through an MCP surface your agents can call.

Install

Mint an ephemeral install URL from the control-plane:

admin$  deploy-gate issue vision-unit --ttl=15m --max-uses=1
  → https://install.wispayr.online/t/<token>

target$ curl -fsSL https://install.wispayr.online/t/<token> | sudo bash -s -- <UNIT_ID>

Installer clones upstream Alice, drops the Protect→Frigate adapter + MCP server overlay, writes alice.conf, starts two pm2 apps, and registers the unit with the control-plane siphon. Re-runs are idempotent.

Spec

Ports4080 (web, SSH-tunnel only) · 4081 (MCP streamable-http)
RuntimePython 3.10+ · pm2 supervisor · systemd cron
Min hardware4 GB RAM · 4 CPU · 20 GB disk · GPU optional
Modelsteacher yolo11x · student yolo11n · ONNX FP16 export
IngestProtect snapshots → Frigate SQLite (adapter, idempotent)
Telemetrysiphon: alice_vision_unit · prism: vision_training_backlog
LicenseAlice CC BY-NC 4.0 · internal/site use only

Pipeline

1. Ingest

Protect events → clean WebP snapshots + Frigate SQLite entries.

2. Dedup

pHash + IoU-box + NMS — kills redundant frames before training.

3. Annotate

YOLO11x teacher auto-labels missing frames at 0.7 confidence.

4. Train

Per-site fine-tune of an 11n student — 30 epochs at 640px.

5. Export

ONNX FP16 drops straight into iris / halio / traffic-vision.

6. Curate

Human-in-the-loop viewer catches edge cases for the next round.

MCP surface (18 tools)

statuscameras_listlive_list datasets_listmodels_listdataset_stats missing_labelsfind_duplicatesadapter_ingest trainer_statustrainer_logsset_dataset pipeline_exportpipeline_deduppipeline_annotate pipeline_trainpipeline_onnxpipeline_run_all

Streamable-http transport at http://<host>:4081/mcp. Drop the URL into any MCP-aware agent and drive the full curation loop without touching the UI.

Fleet rollup

Every registered unit feeds the vision_training_backlog prism lens on big-server. Rollup exposes total cameras, live events, missing labels, active trainings, and the top-10 backlog queue — ready for the Overwatch operator fabric.