Matisse.ai dental shade matching

This is a founder track record project, delivered by our principal’s team at Metaprise Systems between 2019 and 2021, before Lakeshore Labs existed. It is included here because the engineering problems it solved, calibrated capture, color science, and model-driven recommendation, are the same ones we take on today.

The challenge

When a dentist orders a ceramic crown or veneer, someone has to decide what color it should be. The traditional method is a technician holding plastic shade tabs next to the patient’s tooth and judging by eye. Lighting conditions, individual color perception, and plain human error make the result inconsistent, and conventional shade tabs do not have the granularity of natural teeth, which vary continuously in translucency and hue. The industry consequence was stark: up to 30% of ceramic restorations needed costly remakes because the color did not match. Worse, whatever measurement did happen chairside was usually lost by the time the case reached the lab, so the technician mixing the ceramic worked from a shorthand code and a guess.

What we built

The team built Matisse.ai (the name stands for Matching Any Tooth In Shade So Easily), a SaaS platform that turns shade selection into a measurement problem and ceramic formulation into a computation.

Capture

Two input paths feed the system. A native Swift iOS app guides clinicians through calibrated tooth photography chairside, controlling for the capture conditions that make casual photos useless for color work. For practices with dedicated hardware, the platform integrates directly with digital spectrophotometers such as Optishade over a device API, ingesting spectral measurements alongside RGB images.

Color science and the CV engine

A computer vision engine, built on TensorFlow and OpenCV, extracts L*a*b* color coordinates from tooth images using deep models trained on a dental shade dataset. Working in CIELAB rather than RGB matters: it is a perceptually uniform color space, so distances between coordinates correspond to differences a human eye actually sees. That is what lets the system express a target shade as numbers a downstream process can optimize against, and it is how the engine reached 95% color matching accuracy.

Recipe engine

The target L*a*b* coordinates feed a recipe engine that computes, per restoration material, the staining formula, dentin mixing ratios, and micro-layering guidance for monolithic cases. Ceramic color shifts during firing, so the engine also handles post-bake correction: the technician re-measures the fired piece, and the system computes an adjustment recipe, closing a delta-E minimization loop between target and result.

Delivery surfaces

Clinicians get a web product with shade reports and case sync, so the measurement taken at the chair travels with the case. Dental labs receive a ceramic recipe card with layering instructions, replacing the shorthand shade code with an actionable formulation. The backend runs Python and Node.js services with PostgreSQL on AWS, deployed on Kubernetes.

How it was delivered

The engagement ran 8 months with a team of four, in three broad phases. The first months went to the hard part: assembling and annotating the dental shade training data, building the color extraction models, and proving the spectrophotometer integration. The middle phase built the user-facing products, the React web application for lab workflows and the Swift iOS app for chairside capture, with synchronization between them. The final stretch was real-world hardening: beta use in working labs, tuning the models against actual cases, and preparing the platform for production operation on Kubernetes.

What shipped

• A production SaaS platform, live today at matisse.ai, with the iOS capture app on the App Store and 10K+ downloads
• A computer vision engine extracting L*a*b* coordinates from calibrated tooth images at 95% color matching accuracy
• Optishade spectrophotometer integration over a device API
• A recipe engine producing staining formulas, dentin mix ratios, micro-layering guides, and post-bake corrections per restoration material
• Clinician web product with shade reports and case sync, plus lab-facing ceramic recipe cards
• 85% faster shade selection and remakes down 70% for participating labs

The product outlived the engagement, which is the outcome that matters: a measurement pipeline and recommendation engine sound enough to keep running as a commercial product years later.