SITUATION​
Our client developed an iPhone-compatible lens to recommend matching foundation shades by scanning customers' faces in-store using the customer’s mobile app. The device collected image data from key facial regions, processed it through the app, and sent it to a REST API for storage and model training. The recommendation engine used a Random Forest Regressor Machine Learning (ML) model to improve predictions using skin tone data. However, as usage increased, the model's recommendation accuracy decreased, eroding customer and employee confidence and leading to most stores withdrawing the lens device.​

SOLUTION​
Everforth Apex was engaged to revamp the ML-driven recommendation engine. A multidisciplinary team of AI/ML engineers, data scientists, and solution architects collaborated with our client’s stakeholders to deliver a transformative solution.​

Our team began with an in-depth audit of the existing Random Forest model, identifying flaws in how color values were interpreted and weighted.​

In response, we used mathematical color science and introduced advanced color space conversions (e.g., from RGB to CIELAB) to more accurately reflect how human eyes perceive color differences.​

With the groundwork laid by the audit, our team began upgrading the existing ML model with the following changes:​

• Replaced the legacy model with a hybrid ML approach using gradient boosting techniques and neural networks for complex non-linear mappings between skin scan data and product tones.

• Implemented a feedback-driven mechanism that learned from in-store consultant validations and user satisfaction metrics to further refine the model’s recommendations over time.

• Containerized the enhanced ML model as a hardened microservice, deployed within our client’s internal network.

• Delivered robust quantitative validation (color match accuracy metrics, reduced mismatch rates) and qualitative feedback (surveys and A/B testing in pilot stores) to demonstrate improvements.

RESULT​
The revamped ML-driven skin tone recommendation model significantly improved key performance metrics, with the added benefit of being more secure.​

• There was a 65% overall improvement in color matching accuracy​.

• Mobile app usage increased by 28% post-re-launch in pilot stores, indicating renewed customer engagement​.

• The lens technology was reinstated in select markets, leading to higher in-store traffic and product conversions​.

• Documented revenue lift from improved customer trust in recommendations and reduced product returns​.

• Enhanced data security & compliance with a securely hosted ML pipeline within customer’s existing infrastructure.