Production Case Study · Computer Vision & Privacy
TICNES Photo Vault: On-Device Multimodal Embeddings & Zero-Knowledge Search
Impact Summary: TICNES Photo Vault enables individuals, field engineers, and media teams to locate any photograph across tens of thousands of untagged files using natural-language descriptions in Portuguese or English — achieving < 180ms semantic search latency and 96.4% top-5 retrieval accuracy while keeping raw media indexed locally on the user’s hardware.
Contents
1. The challenge: Untagged media sprawl vs. cloud privacy risks
Both personal photo libraries and corporate inspection archives suffer from the same structural bottleneck: cameras and smartphones generate thousands of files with opaque sequential filenames (`DSC_0942.JPG`, `WhatsApp Image 2026-04-12.jpeg`), and almost nobody has the time to tag them manually. Six months later, locating a specific site defect photograph, signed delivery receipt, or family memory requires scrolling through thousands of thumbnails by hand.
Existing cloud photo platforms offer visual search only if the user agrees to upload their entire unencrypted archive to third-party servers. For engineering firms, legal teams, medical practitioners, and privacy-conscious users, surrendering custody of sensitive imagery is unacceptable.
2. The solution: Hybrid local-first indexing and semantic retrieval
We built TICNES Photo Vault around a strict sovereignty guarantee: your images stay on your machine unless you explicitly choose to export an encrypted backup. By running quantized vision-language models directly on the user’s workstation or laptop, Photo Vault builds a rich semantic index locally:
- Natural-language conceptual search: Users can type descriptive phrases in Portuguese or English — such as “quadro elétrico trifásico na obra do Lubango” or “family dinner with birthday cake outdoors” — and retrieve matching photos even when zero manual tags exist.
- Automatic synthetic captioning & editable tags: During idle background indexing, an on-device vision captioner generates descriptive summaries and suggested category tags that the user can refine or lock.
- EXIF & folder-scoped filtering: Users can scope any semantic search to specific directory trees, date ranges, camera models, or GPS coordinates.
3. Inside the multimodal ingestion and query pipeline
As detailed in our companion engineering note on Local-First Computer Vision, Photo Vault combines three complementary retrieval signals inside a unified local database:
| Retrieval Signal | Index Structure | Example Query Handled |
|---|---|---|
| Visual Scene & Object Semantics | HNSW vector graph (cosine similarity over INT8 vision embeddings) | “Excavator near concrete foundation” |
| Embedded Text (On-Device OCR) | SQLite FTS5 trigram index over extracted image text | “Factura n.º 2026/084” or license plate numbers |
| EXIF, Caption & Tag Metadata | B-Tree + FTS5 index over timestamps, camera specs and user tags | “Taken in Benguela between May and July” |
4. Performance and retrieval accuracy benchmarks
We evaluated Photo Vault against a benchmark archive of `50,000` mixed-resolution photographs (construction inspections, scanned documents, event photography, and mobile snapshots) queried with `400` bilingual Portuguese and English search prompts:
- < 180ms end-to-end query latency (p95): Embedding the user’s text query, executing the HNSW nearest-neighbour search, merging FTS5 OCR hits, and rendering thumbnails completes in under a fifth of a second.
- 96.4% Recall@5 accuracy: In `96.4%` of benchmark queries, the target photograph appeared within the top five returned results without any prior manual tagging.
- AES-GCM-256 zero-knowledge encryption: When optional encrypted vault export is enabled, all index shards and media blobs are encrypted client-side before leaving the device.
5. Enterprise and consumer deployment scenarios
Today, Photo Vault serves both individual users managing personal media libraries at photo-vault.ticnes.ai and enterprise teams who need air-gapped or on-premise visual search over field-engineering inspections, insurance claims, and physical asset inventories. To discuss an on-premise deployment for your organisation, contact TICNES AI.