This is a submission for the Hacktoberfest Weekend Challenge: Build for a Friend
What I Built
My mother has cooked incredible traditional meals for our family for decades, but her recipes have never existed on paper. Like millions of home cooks, her kitchen runs entirely on intuition and sensory memory: "fry until the aroma fills the room," "a gooseberry-sized ball of tamarind," and "a small fistful of roasted jeera."
Whenever anyone asks her for a recipe, it's impossible to follow or replicate accurately. On top of that, we've recently been focusing on everyday family wellness at home—trying to incorporate nutrient-dense ancient millets (like foxtail, barnyard, and little millet) and lighter cooking techniques without turning beloved comfort food into sterile "diet food" or ruining the sacred tadka. But every time we searched mainstream apps for "healthy Indian recipes," the suggestions were culturally clueless: swapping lentils for cold kale salads, stripping away mustard seeds, or replacing regional spices with bland Western substitutes.
I built RasoiVault for her.
RasoiVault is an open-source heritage culinary engine. It takes oral cooking fragments, unmeasured notes, and festival recipes, standardizes them into structured measurements, and adapts them for everyday wellness (substituting refined grains with ancient millets and optimizing cooking oils) while strictly preserving traditional tempering order, aroma cues, and mouthfeel. It also features a permanent Family Vault so her recipes are digitized, preserved, and searchable on her kitchen tablet.
Demo
- Live Web Application: https://rasoivault-web.onrender.com
- FastAPI Backend: https://rasoivault-api.onrender.com/api/health
Code
The entire codebase is open-source under the MIT License:
DracFiendMG
/
rasoivault
Family recipe ingestion and loved one's health profile dashboard
🍲 RasoiVault
An open-source heritage culinary engine built for family kitchen sovereignty, powered by Google Gemma 3 and Backboard.io.
Built for the Hacktoberfest Weekend Challenge: Build for a Friend.
đź’ˇ The Core Problem
Traditional home cooking runs on oral shorthand and sensory memories ("fry till aroma changes", "a gooseberry of tamarind"), making cherished recipes impossible to pass down or scale. Mainstream recipe apps provide culturally tone-deaf Western substitutes (swapping lentils for raw kale salads or removing essential tadkas).
RasoiVault bridges traditional intuition with everyday wellness:
- Standardizes messy oral notes into exact metric recipe cards.
- Intelligently swaps refined grains for ancient millets (foxtail, barnyard, kodo) while preserving tempering and texture.
- Maintains persistent household culinary preferences using Backboard.io.
- Archives every adapted dish in a sovereign Family Vault.
🛠️ Tech Stack
-
AI Core: Google Gemma 3 12B (
google/gemma-3-12b-it) via OpenRouter - Memory & Orchestration: Backboard.io (Assistant &…
Repository: https://github.com/DracFiendMG/rasoivault
Architecture:
-
/backend: FastAPI microservice enforcing Pydantic validation schemas, dynamic model configuration, and persistent Backboard memory sessions. -
/frontend: Responsive React + Vite + Tailwind dashboard with side-by-side substitution diffs, wellness tagging, and the persistent Family Vault.
How I Built It
RasoiVault is built around open-weight intelligence and sovereign infrastructure:
Open-Weight Core (Google Gemma 3 12B):
We leverage Gemma 3 12B via OpenRouter as our culinary ethnographer and nutritional transformation brain. Open-weight Gemma was prompted with deep culinary domain constraints—treating traditional Indian tempering sequences (tadka / talimpu) and roasting methods as vital cultural assets while calculating proper grain hydration and starch ratios.Persistent Context Layer (Backboard.io):
Family pantry preferences shouldn't need re-entering every morning. We integrated Backboard's persistent memory layer to create an active assistant thread. Backboard retains her culinary profile, preferred cold-pressed oils, and regional spice tolerances across conversations without prompt drift.Backend & Architecture (FastAPI + Pydantic v2):
The backend strictly validates responses using Pydantic models to guarantee that every recipe returns structured substitutions, health impact breakdowns, and preserved heritage notes without hallucinated formatting.Production Runtime (Render):
Both the FastAPI web service and the React static site are hosted on Render, communicating over an automated environment pipeline with zero downtime.
Why Does Open Innovation Matter?
When building tools for family heritage and daily life, proprietary closed APIs present real trade-offs:
Preserving Heritage Without Cultural Sanitization:
Most commercial chatbots are tuned toward Western dietary paradigms. When prompted with traditional regional dishes, they tend to strip out authentic spices or suggest irrelevant Western staples like quinoa and chia seeds. Open weights like Gemma allow us to instruct the model without corporate guardrail bias eroding authentic culinary traditions.Data Sovereignty for Family Heirlooms:
Decades-old family recipes and personal dietary habits belong to the family, not corporate model-scraping pipelines. Open-source architectures let us keep our data sovereign and self-contained.No Subscription Fatigue for Loved Ones:
Tools built for parents and relatives shouldn't stop working because an API credit limit was hit or an expired card locked them out. With open models, this engine can be hosted permanently on sovereign infrastructure or run locally on a home device using Ollama.
Prize Categories
- Overall Challenge: Weekend Challenge: Build for a Friend
- Best Use of Render: Full-stack deployment on Render (FastAPI web service + Vite static frontend) with automated environment variable linking.
- Best Use of Gemma: Powered by Google's open-weight Gemma 3 12B for low-latency culinary ethnography and nutritional adaptation.
- Best Use of Backboard: Integrated Backboard's Assistant and Thread management with memory persistence to maintain long-term household culinary profiles.
Top comments (0)