Multimodal Asset Recovery AI
Visual + multilingual intelligence for lost & found and urban asset recovery.
CLIP ViT-B/32 | SigLIP2 So400M
The Urban Intelligence Platform uses multiple multimodal AI models, including CLIP and SigLIP2, to convert visual and textual evidence into model-specific representations, retrieve and rank relevant lost-and-found and asset-recovery candidates, and enrich results with OCR, barcode, and contextual signals. CLIP and SigLIP2 operate in separate embedding spaces -- their retrieval evidence feeds the Urban Intelligence workflow, while final recovery decisions remain subject to governed backend rules and human verification.
Live AI Demo
Analyzing multimodal evidence…
CLIP ViT-B/32
—
—
SigLIP2 So400M Experimental / Uncalibrated
—
—
Ranked candidates below are retrieval evidence (OCR/barcode overlap, when available) -- not a confirmed match.
Similarity values are retrieval signals and are not ownership probabilities. Scores from different AI models are not directly comparable.
CLIP ViT-B/32
- Model
- openai/clip-vit-base-patch32
- Embedding
- 512-D
- Role
- Production Baseline
- Status
- Unavailable
SigLIP2 So400M
- Model
- google/siglip2-so400m-patch14-384
- Embedding
- 1152-D
- Role
- Advanced Multimodal Evaluation Engine
- Status
- Unavailable
- Calibration
- Experimental / Uncalibrated
Service Status
- Model
- —
- OCR
- —
- Storage
- —
- Device
- —
Live values from the public readiness endpoint (/health/ready).
Model Provenance
- Model ID
- openai/clip-vit-base-patch32
- Model revision
- 3d74acf9a28c67741b2f4f2ea7635f0aaf6f0268
- Embedding dimension
- 512
- Preprocessing version
- 76fbbe72ea2cc72c
- Scoring version
- clip-match-v1
- Service version
- 07727a5e5c2af38c44037b8f1f73fa5a9726eddc
Model provenance is captured to ensure matching results can be traced to the AI configuration used during inference.
AI Pipeline
- Image / Text
- Preprocessing
- CLIP / SigLIP2
- Model-specific embeddings
- Tenant/site isolated retrieval
- OCR + Barcode + contextual signals
- Ranked candidates
- Urban Intelligence fusion
- Human verification
CLIP 512-D and SigLIP2 1152-D embeddings are maintained and searched in separate model-specific vector spaces.
AI Model Comparison
Live side-by-side comparison is not enabled in this public demonstration console. Protected CLIP/SigLIP2 operations require a signed internal service identity that can never be issued to a browser. Ask the Urban Intelligence Backend team about a demo-safe server-side facade for authorized demonstrators.
CLIP ViT-B/32 (clip_v1)
Latency
Unavailable
Top candidate
Unavailable
Similarity
Unavailable
Model status
Unavailable
SigLIP2 So400M (siglip2_v1)
Latency
Unavailable
Top candidate
Unavailable
Similarity
Unavailable
Model status
Unavailable
AI-assisted candidate retrieval — human verification required.
SigLIP2 is an experimental, uncalibrated engine: its similarity scores are not directly comparable to CLIP's scores or to each other, and neither engine's score is an ownership probability. A higher number never means a more certain match.
Trusted AI
- ✓ Tenant isolated
- ✓ Site scoped
- ✓ Signed service authentication
- ✓ Request traceability
- ✓ Model provenance
- ✓ Synthetic demo-data separation
- ✓ No raw embeddings exposed
- ✓ Human-in-the-loop decision
- ✓ No PII required for model demonstration
- ✓ Production data inaccessible from demo facade