| --- |
| title: Chhaya — Shade For Those Who Work In The Sun |
| emoji: ⛱️ |
| colorFrom: yellow |
| colorTo: red |
| sdk: gradio |
| sdk_version: 5.49.1 |
| app_file: app.py |
| pinned: true |
| license: apache-2.0 |
| short_description: MedGemma skin & heat-health companion for outdoor workers |
| tags: |
| - track:backyard |
| - achievement:tiny-titan |
| - achievement:well-tuned |
| - achievement:sharing |
| - achievement:offgrid |
| - achievement:off-brand |
| - achievement:best-demo |
| - sponsor:modal |
| - build-small-hackathon |
| - backyard ai |
| - backyard-ai |
| - tiny titan |
| - tiny-titan |
| - well tuned |
| - well-tuned |
| - sharing is caring |
| - sharing-is-caring |
| - off the grid |
| - off-the-grid |
| - best demo |
| - best-demo |
| - off brand |
| - off-brand |
| - modal |
| - best use of modal |
| - best-use-of-modal |
| - medgemma |
| - kokoro |
| - healthcare |
| - gradio |
| - zerogpu |
| models: |
| - google/medgemma-1.5-4b-it |
| - build-small-hackathon/chhaya-medgemma-lora |
| - hexgrad/Kokoro-82M |
| datasets: |
| - build-small-hackathon/chhaya-skin-extract |
| --- |
| |
| # Chhaya (छाया — "shade") ⛱️ |
|
|
| **A skin & heat-health companion for the people who work in the sun** — drivers, delivery riders, construction workers, street vendors, farmers. |
|
|
| Heat waves are making outdoor work harsher every year, and the people most exposed to sun damage are the least likely to ever get a skin check. Chhaya is a small, free tool for exactly those hours. |
|
|
| **Built for the Hugging Face Build Small Hackathon (Backyard AI track).** |
|
|
| ▶️ **Demo videos:** [Full walkthrough (4.5 min)](https://youtu.be/-4Fh8WYlUbo) · [90-second cut](https://youtu.be/0465erKuv7Y) |
| 📝 **Blog:** [Chhaya: teaching a 4B model to be the eyes, not the doctor →](https://huggingface.co/blog/build-small-hackathon/chhaya-blog) |
| 📄 [Launch post on LinkedIn →](https://www.linkedin.com/posts/manjunathan-r-06396b1b7_buildsmallhackathon-medgemma-healthtech-share-7472213562757734400-cZby/) |
|
|
| ## What it does |
|
|
| 1. **☀️ Skin check** — snap a photo of any spot or patch of skin. **MedGemma-1.5-4B** describes what it sees (type, colour, borders, symmetry, texture), flags how concerning it looks (`looks ordinary` / `worth watching` / `show a doctor`), and spots visible sun & heat damage. |
| 2. **Plans built around *your* work day** — tell Chhaya your occupation, hours in direct sun, and water access; it builds a hydration schedule and protection plan that fits (drivers learn about window-side UV, riders about helmet-line burns, construction workers about NDMA work-hour guidance). |
| 3. **🌡️ Heat symptom check** — first-aid triage for heat cramps / heat exhaustion / heatstroke. Deliberately **zero AI**: pure deterministic logic from NDMA / WHO first-aid guidance, because emergencies are not the place for sampling. |
| 4. **🗂️ My record** — save checks per body part within the session and re-compare the same spot over time. The spot that *changes* is the spot that matters. |
| 5. **🔊 Read it aloud** — every verdict and the heat first-aid steps can be spoken with one tap (Kokoro-82M, on CPU), in **English or Hindi**, so the plan reaches workers who don't read English easily. |
| 6. **👆 One-tap sample workers** — preloaded persona cards (a bike-taxi rider, an auto driver, a vendor, a farmer) prefill a real skin photo + work-day context and run the check instantly, so anyone can try Chhaya without uploading. Sample photos are open-license images from ISIC + Google SCIN; names/jobs are fictional. |
|
|
| ## Design principle: the model is the eyes, the guidelines are the medicine |
|
|
| MedGemma reads the photo and returns a structured description — that's all it's trusted with. Every medical number and action (litres per shift, ORS timing, 12–3 pm shade window, when to call 108/112) is curated, deterministic content from: |
|
|
| - **NDMA India** heat-action / heat-illness first-aid guidance |
| - **WHO** heat-health advice |
| - **Cancer Council ABCDE** skin self-check criteria (Asymmetry, Border, Colour, Diameter, Evolving) |
|
|
| The model never invents medical advice. Chhaya describes — it never diagnoses. |
|
|
| ## Stack |
|
|
| - **Vision model:** [`google/medgemma-1.5-4b-it`](https://huggingface.co/google/medgemma-1.5-4b-it) (4B, image-text-to-text, bf16) — well under the 32B cap — served with our **fine-tuned LoRA adapter** [`build-small-hackathon/chhaya-medgemma-lora`](https://huggingface.co/build-small-hackathon/chhaya-medgemma-lora) |
| - **Speech:** [`hexgrad/Kokoro-82M`](https://huggingface.co/hexgrad/Kokoro-82M) text-to-speech, run on **CPU** so read-aloud never touches the GPU quota |
| - **Training data:** [`build-small-hackathon/chhaya-skin-extract`](https://huggingface.co/datasets/build-small-hackathon/chhaya-skin-extract) — built from ISIC-2024 (biopsy-anchored) + Google SCIN (phone photos, Monk skin tones 1–7) |
| - **Fine-tune compute:** QLoRA on **Modal** (~$6 of credits, v1 + v2) |
| - **Runtime:** Gradio on Hugging Face Spaces **ZeroGPU** |
| - **UI:** fully custom "step into the shade" theme — full-bleed illustrated hero, striped awning, editorial Fraunces headline, IBM Plex Mono kickers, rubber-stamp verdicts, tactile icon tiles and a dawn-to-dusk sun slider; hero + empty-state art generated with Nano Banana (Gemini image gen) |
| - **Fallback:** with no GPU/token the app runs in demo mode (image heuristics) so it works anywhere |
|
|
| ## The fine-tune: from "describes skin" to "knows when to worry" |
|
|
| The one medically-loaded field — `concern` — is anchored to hard labels, not a bigger model's opinion: ISIC biopsy ground truth + dermatologist-labelled SCIN phone photos. `low` is downsampled and an ABCDE backstop only ever *raises* concern, so the tool can't learn to default to "safe". |
|
|
| | metric (141-image held-out test) | base MedGemma | **Chhaya-tuned (v2)** | |
| |---|---|---| |
| | Valid JSON | 99.3% | **100%** | |
| | Concern accuracy | 33.3% | **69.5%** | |
| | Malignant recall | 0.94* | 0.83 | |
| | Output tokens / answer | 770 | **156** | |
|
|
| \*Base's recall is hollow — it reaches it by dumping ~60% of cases into "watch" (hence 33% accuracy). The tuned model commits to a real triage and runs 5× shorter. Full field notes: [`writeup/BLOG.md`](writeup/BLOG.md). |
| |
| ## Run locally |
| |
| ```bash |
| pip install -r requirements.txt |
| python app.py # demo mode (no GPU needed) |
| ``` |
| |
| For real MedGemma inference: accept the [Health AI Developer Foundations terms](https://huggingface.co/google/medgemma-1.5-4b-it), then: |
| |
| ```bash |
| export HF_TOKEN=hf_your_token |
| python app.py # auto-loads MedGemma if a CUDA GPU is available |
| ``` |
| |
| `CHHAYA_DEMO=1` forces demo mode. |
| |
| ## Credits & inspiration |
| |
| - [Sunny](https://github.com/mrdbourke/sunny) by Daniel Bourke — a MedGemma skin tracker for Australians that proved this shape of app works. Chhaya reimagines the idea for outdoor workers facing heat waves. |
| - Google's MedGemma team; ISIC for the open lesion archives that make this field possible. |
| |
| ## Safety |
| |
| Chhaya is **not a medical device** and provides no diagnosis. Photos are processed in memory and never stored server-side; records live only in the browser session. Anything that worries you on your skin is, by itself, reason enough to see a clinician. |
| |
