Technology
An AI that assists, and never decides alone
Visual recognition of textiles is a hard problem: two fabrics that look alike can demand opposite treatments. Our stance is to face that difficulty rather than paper over it.
Our principles
Four engineering commitments
Multimodal vision
State-of-the-art models, guided by domain instructions written with working dry-cleaning professionals.
Explicit confidence
Every prediction carries its certainty. Below the threshold, control returns to the operator.
Useful grouping
We do not predict a composition label but a treatment path — the datum that actually serves.
Sovereignty
Hosted in France, partitioned per brand, with no resale or advertising use of images.
Learning
The shop floor corrects the model
Every operator correction is a signal. They are collected, analysed, and used to adjust instructions and thresholds — your business makes the system better at your business.
- Corrections traced at source, with the photograph and the drop-off context
- Periodic review of misses: what the model gets wrong is documented, not buried
- Confidence thresholds adjustable by fabric family and by brand
- Evaluation against a reference set before anything reaches production
Architecture
Robust by construction
Documented API
An HTTP interface described in OpenAPI: your tools connect without depending on our interface.
Asynchronous processing
Analyses go through a job queue: a rush slows the computation, never the counter.
Service continuity
Replicated database with automatic failover and encrypted off-site backups.
French hosting
Infrastructure operated in France, under European law, with access logging.
Into the detail
Want to dig into the architecture?
We are glad to document our technical choices, current limitations included. Write to us: we answer seriously.