From Visual Guesswork to AI-Powered Plumbing Part Discovery for Caroma
Ambli designed and delivered an AI-powered visual search system for Caroma that lets customers and staff upload a photograph of a plumbing part and instantly discover visually similar products from the catalogue — with augmentation-trained recognition across angles, rotations and lighting, and a validation layer that rejects irrelevant images.
Executive summary
Ambli designed and delivered an AI-powered visual search system for Caroma that lets customers and staff upload a photograph of a plumbing part and instantly discover visually similar products from the catalogue — with augmentation-trained recognition across angles, rotations and lighting, and a validation layer that rejects irrelevant images.
The business situation
- Customers often did not know the product name, SKU, manufacturer, diameter or fitting terminology.
- Store staff had to visually inspect customer photographs and search inventory from memory and experience.
- Visually similar plumbing components were frequently confused, slowing replacement-part discovery.
- The manual identification workflow did not scale as the catalogue expanded.
The cost of inaction
- Lost sales when a suitable replacement existed in inventory but could not be located quickly.
- High dependency on the tacit product knowledge of individual staff members.
- Inconsistent customer experience across stores and channels.
- No way for customers to independently search the catalogue by shape or appearance.
Ambli's strategic response
- Analysed how staff visually identify plumbing parts — shape, openings, curves, threads, symmetry and connectors.
- Built an image-embedding search grounded in the client’s existing product catalogue.
- Solved the real-world accuracy problem by training with image augmentation across angles, rotations, lighting and orientation.
- Added an image-validation guardrail so the system rejects unrelated images instead of returning misleading matches.
The solution
Mobile and web upload interface for customers and store staff, with ranked visually similar results.
Classifier that determines whether an uploaded image contains a supported plumbing component before search runs.
Resize, normalisation, rotation, flip, brightness, contrast and colour augmentation to expand each product’s visual footprint.
Converts each product image into a numerical representation of its visual and structural characteristics.
Cosine / distance-based similarity search across the indexed catalogue with metadata filtering.
Secure connection to the product inventory so results reflect live SKUs, imagery and descriptions.
Architecture overview
Customers and in-store employees uploading photographs from mobile or desktop browsers.
Upload widget, guidance messaging and ranked visual-match results grid.
Rejects irrelevant uploads (unrelated objects) and asks the user for an appropriate photo.
Preprocessing, embedding generation and similarity search across the augmented product index.
Product catalogue, augmented image dataset, secure object storage for uploads.
Model versioning, logging, performance analytics and quality sampling of search results.
Delivery journey
Business outcomes
| Metric | Before | After | Change | Period | Measured by |
|---|---|---|---|---|---|
| Product discovery method | Manual staff inspection of customer photos | Photo upload returns ranked visual matches | New self-serve pathway | Post go-live | Product analytics on visual-search sessions. Illustrative placeholder — replace with verified client data before publishing. |
| Recognition across photo conditions | Accuracy dropped on rotated / re-lit / re-angled photos | Consistent matches across augmented conditions | Meaningful accuracy uplift | Model evaluation windows | Held-out augmented test set with human-labelled ground truth. Illustrative placeholder — replace with verified client data before publishing. |
| Irrelevant image handling | Any uploaded image returned a “closest” plumbing product | Non-plumbing uploads are rejected with guidance | Absurd results eliminated | Continuous | Validation classifier confusion matrix on evaluation set. |
Evidence & proof
“Customers can now search our inventory the way they actually think about it — by showing us the part. (Placeholder — pending client approval.)”
- Computer vision embedding models (CNN / ViT family)
- Image augmentation pipeline (rotation, flip, brightness, contrast, colour, crop)
- Vector similarity search (FAISS / pgvector / Qdrant class)
- Python FastAPI AI services with REST APIs
- Containerised deployment (Docker) and secure object storage
- Model monitoring, versioning and evaluation tooling
- Secure upload endpoints with size and type validation
- Image validation classifier to reject out-of-scope inputs
- Encryption in transit and at rest for uploaded imagery
- Role-based access to admin and monitoring surfaces
- Model versioning with rollback and audit logging
Why Ambli
Related capabilities
Frequently asked questions
What business problem did the system solve?
Customers and staff could not reliably identify replacement plumbing parts using traditional text search because names, SKUs and dimensions were often unknown.
How does the AI find matching products?
Each product image is converted into a numerical visual representation. When a photo is uploaded, the system compares it against the indexed catalogue and ranks the most visually similar products.
Can the system recognise a product photographed from different angles?
Yes. The model was trained with augmentation across angles, rotations, flips, brightness, contrast and colour variations to handle real-world photo conditions.
What happens when someone uploads an unrelated image?
An image-validation layer checks whether the upload appears to contain a supported plumbing component. Unrelated uploads are rejected with guidance instead of returning misleading matches.
Can it connect with an existing inventory system?
Yes. The visual search layer integrates with the client’s product catalogue so results reflect live SKUs, imagery and metadata.
How long did the implementation take?
Approximately 12 weeks from discovery to launch, including dataset augmentation and accuracy hardening.
What technologies power the solution?
Computer vision embedding models, an image augmentation pipeline, vector similarity search, and a FastAPI-based AI service deployed in containers with monitoring.
Can this approach be adapted to another industry?
Yes — any catalogue-driven business where customers identify products by appearance (spare parts, hardware, fashion, industrial components) can adopt the same pattern.