Velqora

Inspiration

People rarely wonder about their skin in the abstract. They wonder right before buying a new serum, right after an unexpected breakout, or when standing in front of a mirror four weeks into a routine asking: "Is this product actually working for my skin, or is it just marketing?"

Most consumer skincare tools fail this moment in two ways:

Single-snapshot gimmicks: They give you a one-time selfie score with no longitudinal memory, no routine context, and no way to compare baseline versus follow-up measurements over time.
Hallucinated product advice: Generic AI wrappers invent ingredient lists, ignore manufacturer skin-type suitability, or recommend strong exfoliants and retinoids without checking for ingredient conflicts or sensitive-skin constraints.

We built Velqora (https://velqora-0ygv.onrender.com) to close the loop between biometric skin telemetry, evidence-verified eCommerce product discovery, and longitudinal routine experimentation, powered by Perfect Corp.’s YouCam AI Skin Analysis v2.1 and YouCam AI Fitzpatrick Skin Tone v1.0 APIs.

What it does

Velqora is a cyber-clinical AI skin telemetry and eCommerce discovery platform featuring both Dark HUD and Light Lab modes, plus an instant Synthetic Demo (https://velqora-0ygv.onrender.com/demo):

Multi-Modal Biometric Skin Analysis (YouCam API v2.1 + Fitzpatrick v1.0)
Live Camera HUD and Smart Upload: Browser-local front and rear camera capture with a biometric face-framing reticle, plus JPEG, PNG, and server-side HEIC/HEIF WASM decoding.
Automatic Offline Face Framing: Uses local BlazeFace detection to center and crop excess background toward an optimal face-width ratio without upscaling or synthesizing pixels before submitting to YouCam.
15 Skin Indicators and 3 Regional Classifications: Tracks moisture, oiliness, pore, acne, redness, radiance, texture, firmness, wrinkle, age_spot, dark_circle_v2, eye_bag, tear_trough, droopy_upper_eyelid, and droopy_lower_eyelid, alongside whole-face, T-zone, and U-zone skin types.
Optional Fitzpatrick Skin Tone Classification (I–VI): Reuses retained scan photos (with separate explicit per-scan consent) via YouCam’s two-phase face pre-processing and Fitzpatrick v1.0 classification pipeline.
Scan-First Personalized eCommerce Discovery (For my scan)
100-Product Verified Indian Skincare Catalog: Curated from 5 official brand stores (Minimalist, Dot and Key, Deconstruct, Plum, and Re’equil) with SHA-256 provenance hashes, exact variant INR pricing, and 54 verbatim full INCI formulas.
Agentic Live Web Product Discovery (Google Gemini + Search Grounding): Searches across the web for individual Indian skincare product pages matching your derived scan brief, opens candidate URLs, and extracts verbatim price, full INCI, and manufacturer skin-type suitability quotes without trusting unverified LLM claims.
Natural-Language Catalog Filter Parser: Translates conversational shopping requests (such as a fragrance-free moisturizer for dry skin under INR 600 without niacinamide) into a strict, reviewable filter schema while preserving all safety exclusions.
Deterministic Smart AM/PM Routine Builder and Ingredient Compatibility Matrix
Synthesizes a conflict-free Morning (AM) and Evening (PM) routine from your eligible catalog matches (Cleanser to Toner to Targeted Serum to Eye Care to Moisturizer to Broad-Spectrum SPF).
Automatically separates conflicting actives (such as Retinoids versus AHA/BHA exfoliating acids or Benzoyl Peroxide) into different routines or alternating nights, and evaluates multi-product comparisons for conflicts, cautions, and barrier synergies (such as Ceramides plus Hyaluronic Acid, or Vitamin C plus UV Filters).
Experiment Lab and Longitudinal Telemetry (/experiments and /progress/compare)
Clicking "Track This Product" on any catalog or web discovery item pre-fills a baseline-linked skincare experiment.
Users log daily product usage and compare any two completed scans side-by-side. Score deltas are computed deterministically for each metric.
Grounded AI Insights Studio (/insights)
An interactive skincare education and telemetry studio that explains your scan trends and active experiments using American Academy of Dermatology (AAD) reference guidance without ever receiving your raw photographs or making medical diagnoses.

How we built it

Frontend and UI Architecture: Built with Next.js 16.4 (App Router), React 19, TypeScript 7, Tailwind CSS 4, Framer Motion, Recharts, and custom CSS variables powering a dual Cyber-Clinical Dark HUD and Light Lab glassmorphic design system.
YouCam API Integration (lib/youcam/):
Engineered a strict server-only client for YouCam AI Skin Analysis v2.1 and YouCam AI Fitzpatrick Skin Tone v1.0.
Implemented direct YouCam File API pre-signed PUT uploads, Sharp EXIF orientation and metadata stripping, heic-decode WASM container parsing, BlazeFace offline face framing, Zod response validation, idempotent scan submission keys, and atomic PostgreSQL polling claims.
Hybrid Agentic and Deterministic eCommerce Engine (lib/products/ and lib/ai/):
Combined a deterministic rule engine with Google Gemini 2.5 Flash for structured filter parsing and Google Search grounded live web discovery.
Weighted product ranking is transparent and deterministic based on matched categories, ingredients, INCI completeness, skin-type suitability, and preferred brands.
Database, Auth and Privacy (supabase/migrations/):
Supabase PostgreSQL, Auth, and Private Storage (skin-photos): Enforced Row-Level Security (RLS) across all tables (profiles, consent_records, skin_scans, skin_scan_metrics, skin_type_results, skincare_experiments, product_usage_logs, ai_insights, products, and user_product_preferences).
Short-lived signed image URLs, configurable photo retention with daily cron cleanup, and one-click full application data deletion.

Challenges we ran into

YouCam File Downloader and Long Signed URLs:
Early in testing, passing long Supabase signed URLs with query parameters directly to the YouCam task endpoint caused provider downloader rejections because the query string was treated as part of the filename. We solved this by integrating the YouCam File API first—registering the stripped JPEG, uploading bytes via the provider's pre-signed target, and submitting the resulting file identifier to both endpoints.
Preventing AI Hallucinations in eCommerce Recommendations:
LLMs frequently guess ingredient lists or assume a product is suitable for dry skin without proof. Furthermore, live web searches often return category grids or retailer homepages instead of individual products.
We solved this by building a two-stage deterministic verification pipeline: first rejecting category grids and bare domain roots, and second verifying every extracted quote verbatim against normalized source page text before qualification.
Conservative Safety for Structural Concerns and Active Ingredients:
Indicators like tear trough, eye bag, and droopy upper eyelid are structural, and topical creams cannot lift or fill them. Similarly, a scan's cosmetic skin age estimate must never be used to infer legal adult age for strong retinoids. We encoded strict rule boundaries so structural eye indicators return honest educational guidance rather than false promises.

Accomplishments that we're proud of

True Dual YouCam API Integration: Seamlessly combining YouCam AI Skin Analysis v2.1 and YouCam AI Fitzpatrick Skin Tone v1.0 in a single privacy-first workflow that reuses retained photos without redundant uploads.
Zero-Hallucination eCommerce Discovery: Building an eCommerce engine where every single recommendation is backed by verifiable ingredient, suitability, and price evidence, plus automatic basic-care fallback.
Deterministic Smart Routine and Conflict Checker: Going beyond single-product cards to automatically assemble conflict-free AM/PM routines and flag active-ingredient clashes before purchase.
149 Automated Tests and Cyber-Clinical UI: Achieving 149 passing tests across 25 test suites using Vitest alongside a responsive biometric interface that works seamlessly on desktop and mobile.

What we learned

Biometric APIs shine when paired with longitudinal state: A single YouCam scan is fascinating, but pairing two scans across a 30-day product experiment turns raw scores into actionable personal discovery.
Hybrid Deterministic and Agentic AI is the future of retail: Using LLMs for intent parsing and live web retrieval while keeping eligibility, ingredient exclusions, and routine conflict checks 100 percent deterministic gives shoppers conversational magic with clinical trust.
Privacy-by-design builds consumer confidence: Separating live camera previews, expiring private photo storage, and numerical-only LLM context makes users comfortable tracking their skin journey over time.

What's next for Velqora

YouCam Makeup and Apparel Virtual Try-On (VTO) Integration: Expanding Velqora’s eCommerce workflow so users can virtually try on tinted sunscreens, foundations matched to their Fitzpatrick classification, and coordinated apparel directly inside the product explorer.
Expanded Multi-Region Verified Catalogs: Growing our SHA-256 source-verified catalog beyond the initial 100 Indian skincare products to 1,000+ global dermocosmetic formulas with automated provenance re-verification.
Barcode and INCI Label OCR Scanner: Letting users snap a photo of a physical product bottle in-store to instantly run Velqora’s deterministic ingredient conflict checker against their latest skin scan and current AM/PM routine.