TwachaAI

A selfie skin check for India: YouCam AI scores your skin, then builds a routine from Indian brands that fits your monthly budget.

Comment

Inspiration

Twacha (त्वचा) is Hindi for skin. In India, skincare advice comes from three places: a cousin, an influencer, or a ₹1,000 dermatologist visit that many people put off. The result is a bathroom shelf of products bought on hype, often wrong for the skin type and the city. Humid Mumbai skin, dry Jaipur skin and Delhi-smog skin need different things. We wanted something that starts from what your skin actually shows, speaks your language, and respects a student's ₹500 a month as much as a ₹3,000 budget.

What it does

  1. You upload a selfie (or use the camera with a face guide), pick a monthly budget (₹500 / ₹1,500 / ₹3,000) and your climate (humid / dry / polluted metro).
  2. The YouCam Skin Analysis API scores 8 concerns: hydration, oil balance, acne, pores, redness, texture, dark circles and spots.
  3. You get a report card: each concern with a score, a Good / Watch / Focus level and a one-line explanation in English or Hindi. Tap a concern to see YouCam's overlay of where it shows on your face.
  4. A deterministic routine engine turns scores into an AM/PM routine by ingredient (salicylic acid for acne, niacinamide for oil and pores, vitamin C or alpha arbutin for spots, ceramides and hyaluronic acid for dryness), with sunscreen SPF 50 PA++++ every morning.
  5. It fills each step with real, widely sold Indian products (Minimalist, The Derma Co, Dot & Key, Cetaphil, Re'equil, Plum, Lakmé and more) and keeps the monthly cost under your budget. Whatever doesn't fit is shown as "add later". Switch budget or climate and the plan updates instantly.
  6. Re-scan in 4 weeks and see per-concern progress, stored privately on your device. One tap adds a calendar reminder.
  7. Clear safety copy throughout: it's a cosmetic check, not medical advice, and very low acne scores prompt a dermatologist visit.

How we built it

  • Backend: FastAPI (Python 3.12). The YouCam V2 flow runs fully server-side so the API key never reaches the browser: POST /s2s/v2.0/file → PUT the bytes to the presigned URL with the returned headers → POST /s2s/v2.0/task/skin-analysis with src_file_id, dst_actions, miniserver_args.enable_mask_overlay and format: "json" → poll GET /s2s/v2.0/task/skin-analysis/{task_id}.
  • Image prep: Pillow fixes EXIF rotation and downscales to ≤ 2560 px, then picks HD actions (hd_moisture, hd_pore…) when the short side is ≥ 1080 px and SD actions (moisture, pore, dark_circle_v2…) otherwise. SD and HD are never mixed.
  • Score handling: we display YouCam's ui_score and rank priorities on raw_score; region-aware HD results (pores by forehead / nose / cheek) are kept as detail.
  • Routine engine: pure Python rules with no LLM, so the same scores, budget and climate always give the same routine. It handles skin-type detection, sensitivity (skip acids when redness is high), climate nudges, actives that shouldn't share a night, and a per-budget cap on serums.
  • Budget fitting: each product has an approximate MRP and how many months it lasts; essentials go in first at the cheapest suitable price, treatments are added by severity while the monthly total fits, and leftover budget upgrades the essentials to better-matched picks.
  • Frontend: static HTML/CSS/JS with no build step, mobile-first, camera capture via getUserMedia, an EN/हिंदी toggle and localStorage history.
  • Quality: a pytest suite covers the routine engine across every budget × climate × skin profile, the YouCam client with mocked HTTP (success, engine errors, 401, bad parameters, upload failure, timeout), the API endpoints, real captured YouCam responses, and a repo-hygiene check that .env is never tracked.

Challenges we ran into

  • Making ₹500 honest. A good SPF 50 PA++++ sunscreen alone costs about ₹270 a month when used generously. Instead of pretending a 6-step routine fits, the engine guarantees the three essentials and shows what to add when the budget allows.
  • Two kinds of score. ui_score is deliberately encouraging, while raw_score is the model's direct prediction. In real responses a hydration raw_score of 31 came back as ui_score 63. Showing one and prioritising on the other kept the UI kind without hiding real concerns.
  • HD vs SD. They can't be mixed and have different minimum resolutions, so the server picks one per photo after resizing.
  • Hindi copy had to be plain and natural, not a literal translation of cosmetic jargon.

Accomplishments that we're proud of

  • An end-to-end YouCam integration verified against the live API, with real HD and SD responses used as test fixtures.
  • A routine engine that is explainable: every product says which concern it's for.
  • A budget promise that is tested for every budget, climate and skin profile.
  • A demo mode built on a real response, so judges can run it locally without a key (pip install -r requirements.txt && uvicorn app.main:app).

What we learned

  • YouCam's async task model (file → upload → task → poll) and its engine error codes, which map neatly to friendly retake tips ("move closer", "find better light").
  • How region-level HD results (pores by forehead, nose, cheek) can make advice more specific.
  • That a deterministic rules engine is easier to trust, test and explain than a generated routine.

What's next

  • Live prices and stock from Indian retailers (Nykaa, Amazon, Tira) and affiliate checkout.
  • YouCam Skin Simulation to preview what four weeks of the routine could look like.
  • More Indian languages (Tamil, Bengali, Marathi) and voice explanations.
  • Opt-in cloud sync for progress, and a "share with my dermatologist" PDF.

Built With

Share this project:

Updates

Submission history