Reflekt: AI Skin Analysis + Virtual Try‑On

Reflekt - expert skin guidance, operated by AI

The problem

A proper skin consultation in India means a dermatologist appointment: a wait, a commute, and ₹800–2,000 before anyone looks at your face. So most people never get one. They guess buying products off influencer lists and marketplace reviews, with no read of their own skin. Expert guidance exists; access to it doesn't.

What Reflekt is

Reflekt turns one selfie into expert-grade skin guidance for the price of a coffee. A user takes a portrait; Perfect Corp's Skin AI scores twelve concerns; and then the Baymax Agent our autonomous diagnostic layer built on Google Gemini does what a consultant does: it names the priority concern, writes a professional assessment, sets a strategy, builds a personalised AM/PM routine, recommends specific products with reasons, and issues a Next Best Action. It ends with a branded, downloadable PDF report.

The expert in the loop is an AI agent. That is the whole point.

How AI operates the business

Reflekt isn't an app that contains AI - the core service is delivered by AI, in production, on every paying transaction:

The Baymax Agent runs server-side on Gemini, constrained to a strict JSON responseSchema that is validated before anything reaches the user. It produces the entire diagnostic assessment, strategy, routine, product rationale, next action with no human writing or reviewing the output.
Every run is logged as evidence [reflekt-baymax] execution lines and an admin endpoint (GET /api/admin/agent-logs) recording the model, the decision, and a non-sensitive input summary (never raw images or PII). That log is the proof that AI is running continuously, not staged.
The economics are enforced in code, not by a person. Baymax costs 2 credits, charged once per analysis, on success only; re-opening a past report serves a cached result for free no credits, no Gemini call. Insufficient balance routes the user to top up. Charging is idempotent.
Guardrails are deterministic. The agent gives cosmetic and styling guidance only; anything clinical is routed to a dermatologist. The AI decides the advice; it never touches money or overrides safety.

Humans (me) do three things the AI doesn't: safety review of the guardrails, the payment and compliance setup, and customer conversations. Everything a consultant would actually say is the agent's job.

Traction

Launched to real users and generating real revenue in the first hours - [₹6,000 in the first ~6 hours from paying customers]. Payments run on Razorpay in INR; every diagnostic is a metered, paid transaction. Reports are archived per user to a GitHub backup repo (reports/<user>/<date>/), so the record of delivered work is auditable.

How it's built

Vanilla JS front end, serverless functions on Vercel, Upstash Redis for state, YouCam (Perfect Corp) for skin analysis, Google Gemini as the diagnostic brain, Razorpay for payments. The Gemini key lives only on the server. The Baymax product catalogue is real Amazon-affiliate items, each paired with a Gemini-written reason and an add-to-cart link, so the recommendation is monetisable and honest at once.

What I learned

The line between "AI-powered" and "AI-operated" is the whole game. Moving the diagnostic from a feature I write to an agent that decides schema-validated, logged, charged, cached is what turns a skincare app into a professional-services business that runs itself. And the unglamorous parts (charge-once idempotency, cached re-opens, evidence logging) are what make that claim survivable under scrutiny.

What's next

More agents across the operation support triage and lifecycle outreach already drafted and a B2B path: embedding the Baymax consultation into salons and clinics, where one partner is worth hundreds of individual users.