How We Built Faceloria's Face Analysis Pipeline

People often assume a face analysis tool is a single giant AI that "looks" at a face and judges it. In practice, the pipeline is a chain of smaller, carefully controlled steps — and the design choices at each one decide whether the result is trustworthy. Here is how we actually built it.

Step 1: Finding the landmarks

The foundation is on-device landmark detection using MediaPipe (Google's machine learning framework for real-time perception). It processes your photo and returns a set of facial landmarks — precise coordinate points marking your eyes, nose, mouth, jaw, and other features.

Landmark models are fast, but their raw output can jitter frame to frame. That is why all of our downstream analysis works from a stabilized landmark set rather than trusting individual detections.

Step 2: A rules engine, not a neural net

Here is a choice that might surprise you: the actual judgments — face shape, ratios, scores, and feature notes — do not come from one black-box model. They come from a rules engine applied to the measured geometry.

Why? Because rules are explainable. When a report says your face reads as oval because of a specific length-to-width ratio, that is a statement we can trace, verify, and correct. A large model that "just knows" is faster to build but impossible to audit. For a product whose entire value is trust in the measurement, explainability wins.

Step 3: Calibrating the metrics

Numbers are only as meaningful as their reference points. We maintain a metric-calibration layer that maps raw measurements onto interpretable values and reasonable reference ranges, and a result-readings system that turns those into plain-language notes. This is what lets the report explain each measurement instead of dumping raw figures on you.

Step 4: Transparency and evidence

Everything is cross-checked by construction:

• The landmark face map draws your detected points directly onto the photo, so you see exactly what was measured.
• The view-model keeps the source geometry and the presentation separate, so the measurement logic and the display logic do not drift apart.
• Any derived value is computed from the same source differently, so it can be verified independently.

Privacy by design

Because the raw material is a photo of your face, privacy shaped the architecture:

• Landmark processing runs on your device where possible, so your image does not have to leave your machine.
• Where analysis requires server-side work, the data is encrypted and isolated.

The design principle is that you should own your analysis and be able to understand it.

What this means for you

The result is a face report you can trust: reproducible, explainable, and honest about what each number describes. It is the same technology behind every face tool on the site — and the new face report shows the landmark map and explained measurements so you can verify the whole thing yourself.