Nouri: Keep what works from home

Good food. Healthier tomorrows. Nouri is a health helper for a whole household, family or flatmates. It makes Mum's recipes healthier, plans the week with one shopping list, checks any food or skin-care label for everyone at home, and fact-checks Nani's remedies. It works in any language, by voice, text or photo.

Built for the Nebius × NVIDIA Global AI Hackathon (Best Apps & Agents) with NVIDIA Nemotron 3 on Nebius Token Factory.

• Live app: https://[your nouri domain].vercel.app
• Demo video: [YouTube link]
• Screenshots: docs/gallery/

What it does

Feature	What a family gets
Household	Set up once in 3 steps: names, ages, several health conditions and goals per person, diets (vegan, Jain, halal, kosher-style…) and allergies. Share with an invite link, no accounts.
Food	Say a recipe the way it's said at home ("1 katori arhar, ghee on top"). Get it lighter for diabetes, blood pressure or cholesterol, with before → after numbers from food tables.
Week + one shopping list	A dietitian-style week around the family's own dishes: calorie targets, goals, diets and allergies, a note to the parent in their language, and one consolidated list.
Scan & check	Photograph any food or skin-care label. Get an answer for each person (fine / now and then / check the label / not for them), an ingredient breakdown, and "add to list".
Skin & hair	Check a home remedy (keep, fine with care, swap or stop, with sources) and get a simple routine per person from a cited ingredient library.
Insights	Daily aims per person (calories, protein, fibre, sodium, sugar, carbs per meal, water, activity), each cited to WHO, ICMR-NIN, ADA or EFSA.

How NVIDIA Nemotron and Nebius Token Factory are used

Each step goes to the Nemotron model sized for its job. All of them are called through Token Factory's OpenAI-compatible API (app/llm.py), so one client serves every model with no GPUs to manage.

Step	Model on Token Factory	Why this size
Read a spoken or typed recipe into grams	Nemotron 3 Nano (30B-A3B), thinking off	Fast, cheap structured extraction (~2 s)
Make dishes lighter, answer "Dad hates cauliflower", suggest a dish	Nemotron 3 Super (120B-A12B), thinking off	Judgement inside strict rules (2–5 s)
Check remedies, care plans, translation, explain scanned labels	Nemotron 3 Super	Same: constrained reasoning, multilingual
Plan the whole week and write to the parent	Nemotron 3 Ultra (550B-A55B), thinking on	Long-horizon planning across many dishes (~45 s); falls back to Super
Copy the words off a label photo	Gemma 3 27B (vision, on Token Factory)	No Nemotron model on Token Factory takes images yet; it only transcribes, and Nemotron plus code decide

Where Token Factory helped:

• One API key and one client for four models, so routing by task is a one-word change.
• Per-token pricing let us budget each step. A full household session is about 20 calls and 70k tokens.
• chat_template_kwargs switches Nemotron's thinking off for the fast steps.

After a scan, an optional Tavily web search (/api/scan/news, on when TAVILY_API_KEY is set) shows up to 3 recent links on recalls or recipe changes for that product. It never affects a verdict, and only the product name is sent.

Every call is traced in LangSmith (model, tokens, latency) when LANGSMITH_TRACING=true.

Trust: the AI never makes up numbers or safety rules

• Nutrition is computed in code from IFCT 2017 and USDA SR28 (data/foods.csv, 205 foods with home-language names). Anything unmatched is shown as "not counted".
• Calorie targets use Mifflin–St Jeor with WHO Asian BMI cut-offs (app/dietitian.py).
• Safety runs in code after the model answers (app/scan.py, app/care.py, app/dietitian.py):
  • allergies, including "may contain", compound words and Hindi/Hinglish names
  • diets, pregnancy, age limits and prescription-only items
  • dangerous remedies and ingredients (steroids, mercury, hydroquinone) always get "stop"
• Guardrails before the model (app/guards.py): self-harm signals get helplines and never reach it. Poisons or chemicals typed as food get a firm "not food" with emergency numbers; swear words, non-label photos and gibberish get a friendly reply.

Run it

git clone https://github.com/ishtrain07/heritage-plate.git && cd heritage-plate
pip install -r requirements.txt
cp .env.example .env               # add NEBIUS_API_KEY for live mode
export $(grep -v '^#' .env | xargs)
uvicorn app.main:app --reload      # http://localhost:8000

• Demo mode (no key, or MOCK=1): deterministic rules stand in for the models, so every screen works offline.
• Live mode: set NEBIUS_API_KEY. python scripts/check_models.py verifies the model IDs. HP_LIVE=1 python scripts/live_run.py runs the demo live and reports latency and tokens.
• Settings:
  • RATE_LIMIT: requests per client per 10 minutes, default 200
  • MAX_CALLS_PER_HOUR: model calls across all users, default 600
  • TRUST_PROXY
  • LLM_TIMEOUT_S, LLM_TIMEOUT_ULTRA_S, LLM_DEADLINE_S
  • VISION_MODEL, default google/gemma-3-27b-it
  • TAVILY_API_KEY (optional): turns on "In the news" after a scan
  • LANGSMITH_TRACING, LANGSMITH_ENDPOINT, LANGSMITH_PROJECT, LANGSMITH_API_KEY

Deploy: Vercel, using vercel.json (app/main.py as a Python function, maxDuration 120 s). Set NEBIUS_API_KEY (and the LangSmith variables) for Production and Preview. A Dockerfile is included for any container host.

Tests

python -m pytest -q    # 272 tests, offline (mock mode)
ruff check .

How we built it with review agents

.claude/agents/ holds the three Claude Code agents that audited every change before it shipped:

• code-reviewer.md: correctness and safety. It found allergy false negatives in the scanner, each now a regression test in tests/test_scan.py.
• frontend-ux.md: design, accessibility, dark mode, phone layouts.
• product-ux-strategist.md: flow, copy, and nutrition guidance from a dietitian's view.

Project layout

app/main.py        routes: /api/parse, /api/adapt, /api/suggest, /api/plan, /api/week, /api/demo, /api/health
app/scan.py        /api/scan: label photo → text → per-person checks (allergies, diet, sugar/salt, skin safety)
app/insights.py    /api/insights: daily aims per person, pure code
app/care.py        /api/care/*: remedy check, care plan, label check, cited library, safety rules
app/dietitian.py   week plan: calorie target, diets, allergies, portions, shopping list
app/engine.py      parse → adapt → plan pipeline (live Nemotron or mock rules)
app/llm.py         Token Factory client, JSON parsing, cache, budget, deadline, Ultra → Super fallback, tracing
app/prompts.py     system prompts
app/guards.py      input guardrails, rate limit, model-error handling
app/nutrition.py   nutrition maths over data/foods.csv
data/              foods.csv, ingredients.csv, meals.json, care_actives.json, demo_families.json, facts.json
static/            the app: index.html, styles.css, family.js, app.js, care.js, scan.js, setup.js, insights.js, shop.js
scripts/           build_foods.py, check_models.py, live_run.py, build_preview.py
docs/              DEVPOST.md, VIDEO_SCRIPT.md, gallery/, notes/ (specs and review notes)
tests/             pytest suite

Rebuilding the food table

npm pack @ifct2017/compositions fda-nutrient-database && for f in *.tgz; do mkdir -p "${f%.tgz}" && tar -xzf "$f" -C "${f%.tgz}"; done
python scripts/build_foods.py --ifct ifct2017-compositions-*/package/index.csv --sr28 fda-nutrient-database-*/package/data/ABBREV.txt

Credits

Photos on Unsplash by Anju Ravindranath, Helena Lopes, Omar Abozeid, iKshana Productions, Vije Vijendranath and Jack Lee. Food data: IFCT 2017 (National Institute of Nutrition, India) and USDA SR28.

Disclaimer

General guidance, not medical advice. Check changes with a doctor, pharmacist or registered dietitian.

License

MIT