I got tired of every appearance AI doing the same thing: upload your face to their servers, get a score, get told to "moisturize more." The problem isn't the advice quality — it's that there's no *plan*. A 7.4/10 rating doesn't tell you what to actually change. And uploading your selfie to a random startup's server to find out you're average felt like a bad trade. So I built MirrRd around two constraints I couldn't compromise on: **1. The photo never leaves your device.** MediaPipe runs the face analysis entirely in-browser — face shape, jaw ratios, skin tone, eye spacing, all extracted client-side. The server only sees a JSON blob of measurements. You can verify this in network traffic. **2. The output has to be specific enough to act on.** Not "get a better haircut" — the exact cut, the barber terminology, what to ask for. Not "dress better" — actual brand names and fits for your detected face shape and build. The celebrity goal input came from watching how people actually talk about appearance goals. Nobody says "I want an oval face" — they say "I want to look like Ryan Gosling." So that became the input model. The approach evolved a lot. Early versions were too generic — GPT would give advice that could apply to anyone. It took about 40 prompt iterations and adding an independent evaluator pass (a second LLM call that grades each result 1–10 against a stylist rubric) before the output was specific enough to be useful. Would genuinely love feedback on whether the output quality holds up — that's the thing I'm still calibrating.

MirrRd,

Brutally Honest Appearance Advice

Brutally Honest Appearance Advice