Product on Trial

Is your skincare actually working? Put it on trial. YouCam Skin AI scans are the evidence, and the court rules: Acquitted, Guilty, or Case Continued. Stop guessing, start judging.

Comment

Inspiration

Everyone has a shelf of skincare they think might be working. Results take weeks, mirrors are biased, and a bad night's sleep looks exactly like a bad serum. So people keep paying for duds, or give up on good products too early.

The hackathon brief described the moment I wanted to build for: standing in front of the mirror, deciding whether a product is working. I turned that question into a court case.

What it does

Product on Trial puts your skincare on trial, and your skin is the witness.

  1. File a case. Name the product and pick what it claims to do (hydrates, clears breakouts, boosts glow…). Each claim has a realistic time-to-results: hydration ~7 days, texture ~28, acne ~42. Nothing gets judged too early.
  2. Submit evidence. Take a selfie at the start and at each check-in. The YouCam Skin Analysis API scores 8 concerns (acne, redness, oiliness, pores, texture, moisture, radiance, fine lines), and YouCam's detection maps appear as "forensic photos".
  3. Witness statements. Log anything else that changed (bad sleep, stress, period, a new product, weather). These become objections that lower the court's confidence, so the product isn't blamed or credited unfairly.
  4. Hear the verdict:
    • ✅ Acquitted: it's working, keep it on the shelf
    • ⛓️ Guilty: it's irritating your skin, or it did nothing after a fair trial. Off to product jail
    • ⏳ Case continued: not enough evidence yet, and the court sets your next hearing date

Every ruling lists the charges, the evidence and concrete next steps (e.g. "patch-test on your jawline", "add a ceramide moisturizer"). The docket tracks how much money you've stopped wasting on duds.

How I built it

  • Next.js 16 + TypeScript + Tailwind CSS, deployed on Vercel.
  • YouCam APIs: the File API (register, then a signed PUT upload), then Skin Analysis V2.1 with 8 SD concerns in JSON format with mask overlays, polled with backoff. Engine errors such as error_lighting_dark or error_src_face_too_small become friendly retake instructions.
  • A deterministic verdict engine, so verdicts are consistent and explainable. For each concern $c$, with baseline score $s_0$, the latest evidence averages the last two scans so one bad photo can't decide a case:

$$\Delta_c = \frac{s_{n-1} + s_n}{2} - s_0$$

Changes within a noise band of $|\Delta_c| < 5$ are ignored. A weekly trend (least-squares slope) checks that changes are consistent. Charges are raised from the evidence:

  • Aggravated Irritation: redness or acne drops by 6+ points
  • Breach of Promise: a claimed concern gets worse
  • False Advertising: a claimed concern barely moves after its expected time
  • Collateral Damage: an unclaimed concern gets worse

The court's confidence combines scan count, time on trial relative to what the claims need, and trend consistency, minus a penalty for recent confounders:

$$\text{confidence} = 0.4\,\min!\left(1, \tfrac{n-1}{4}\right) + 0.4\,\min!\left(1, \tfrac{t}{T_{\text{claims}}}\right) + 0.2\,\text{consistency} - 0.15\,[\text{confounders}]$$

  • Rulings are written in a playful courtroom voice with concern-specific advice, but the numbers always decide the verdict.
  • Privacy: cases and photo thumbnails stay in the browser. Photos are only sent to YouCam for scoring.
  • A soft daily scan cap protects API units on the public deployment.

Challenges I ran into

  • Noisy single scans. Lighting and angle move scores by a few points. Rather than trust any one number, I built in baselines, a noise band, averaging and a confidence score.
  • Being honest with an AI. Generated text can drift from the data, so the verdict comes only from the scan numbers, and the text explains it.
  • Side effects hide. A product can improve what it promises while quietly irritating your skin, so every scan checks all 8 concerns, not just the claimed ones.
  • Demoing a multi-week trial in 3 minutes. I added a demo mode that fast-forwards the calendar while every scan stays real.

Accomplishments that I am proud of

  • A verdict that catches a product that "works" but irritates your skin.
  • Fully explainable rulings: every verdict cites its exact numbers.
  • An experience playful enough that people want to check in: mugshot-style scans, rubber-stamp verdicts and product jail.

What I learned

How YouCam's skin scores behave across sessions, why baselines and realistic timelines matter more than any single scan, and that a clear verdict with next steps is far more useful than a dashboard of numbers.

What's next for Product on Trial

  • Check-in reminders
  • Head-to-head trials ("Serum A v. Serum B")
  • Ingredient-aware advice when a product is convicted
  • Shareable case files
  • Routine-level trials that track several products at once

Built With

Share this project:

Updates

Submission history