webAI · YOLO26-MLX build challenge · 2026

Vault

Document your home in 5 minutes.
Nothing leaves your phone.

The problem

Filing a home-insurance claim means giving a stranger 200+ photos of your house.

Existing inventory apps make it worse: they upload everything to the cloud.

$2,200
avg. underclaim per US household due to incomplete inventories
(III, 2024)

The demo

  • Walkthrough Scan — RoomPlan LiDAR and YOLO inference running simultaneously on a single ARSession, one camera feed, real-time
  • Tap to bulk-price — LLM-typed reports from the on-device catalog
  • Generate PDF — insurance-ready, with 3D layout + evidence crops
  • Rescan later — diff view shows what was added or moved

The 10× moment

Same iPhone. Same room. Two models.

Stock YOLO26-s (COCO 80)

  • sofa
  • chair
  • tv
  • potted plant
  • person × 2
  • microwave
  • kettle
  • photo frame
  • wardrobe
  • air conditioner

[FILL: e.g. 5] items detected

→

Fine-tuned YOLO26-m (household 32)

  • sofa, chair, tv, potted_plant
  • microwave
  • kettle
  • photo_frame × 3
  • wardrobe
  • air_conditioner
  • remote_control
  • rice_cooker
  • fridge

[FILL: e.g. 17] items detected

Pure MLX, end to end

RoomPlan + YOLO on one ARSession — LiDAR mesh and object detection share the same camera feed, both real-time on iPhone. 

# 1. Training (on your Mac, pure MLX, no cloud GPU)
yolo-mlx ← HomeObjects-3K + Roboflow household
       │  3,622 images · 32 classes · 50 epochs
       ▼
best.safetensors  (yolo26-m, ~84 MB)
       │
       ▼
# 2. Drop into iOS app bundle
Vault/Resources/yolo26m.safetensors
       │
       ▼
# 3. Inference on iPhone (pure MLX-Swift)
YOLO26Predictor ── 5 fps live ──┐
                              │      ▼
                              │   RoomPlan (LiDAR)
                              │      │
                              ▼      ▼
                       Catalog + 3D mesh
                              │
                              ▼
       # 4. Only structured text leaves the phone
       LLM ← {item: "MacBook Pro", count: 1, value: $2,000}
       # Photos? Never.

By the numbers

0.74
val mAP50
(best, epoch 45 of 50)
7h
training time
M1 Max, batch 8
84MB
model on phone
(yolo26-m safetensors)
0
photos
uploaded

Stock COCO 80 → fine-tuned household 32. Trained on the user's Mac. Shipped in a 1.5GB IPA. Inference real-time on A18.

The fine-tune in one chart

Training metrics curve

Generated locally on M1 Max from train.log via marketing/gen_chart.py. Re-runnable any time.

What's next

Short term

  • Port yolo26-m architecture to MLX-Swift (currently s; m widths are 0.75)
  • Add ~200 user-contributed images to weak long-tail classes (kettle, hairdryer, etc.)
  • TestFlight + App Store submission

Stretch

  • YOLO-World port to MLX — open-vocab fallback for the "type any item" case
  • Re-scan diff view → automatic insurance addenda
  • Web-app dashboard (encrypted backups only, photos still never leave the phone)
Thank you, @thewebAI

Vault

Built with yolo-mlx + Apple MLX.

@promptforce · github.com/PromptForcePrime/vault-app