From receipt to recipe — with your pantry in the loop.
Aishi is an AI kitchen assistant that connects what you buy, what you have, and what you can cook — with human approval before anything important is saved.
Home cooking breaks down in the gap between three things that rarely talk to each other:
- What you bought (receipts, shopping)
- What you still have (pantry, expiry)
- What you could make (recipes, cravings, leftovers)
Most tools handle one piece: a recipe app, a shopping list, or a pantry spreadsheet. None of them reliably answer:
> "Given what's in my kitchen right now — especially what's about to go bad — what should I cook tonight?"
Aishi is a full-loop kitchen assistant:
1. Chat — Ask in plain language: paste a recipe URL, upload a receipt, or say what you feel like eating.
2. Extract — Pull structured recipes from URLs and grocery lines from receipt photos.
3. Track — Build and maintain a pantry of lots (each purchase is its own item, with dates and storage).
4. Suggest — Recommend recipes from your saved collection using semantic search plus pantry fit and expiry urgency.
5. Cook — Enter cooking mode, confirm ingredient use, and deduct pantry quantities only after you approve.
Nothing sensitive is written to the database until you confirm it.
Less food waste | Suggestions can weight ingredients that expire soon.
Less mental load | Receipt upload updates pantry instead of manual entry.
Better decisions | Rankings blend "matches my craving," "I already have it," and "use it before it spoils."
Trustworthy automation | Recipe saves, receipt commits, and pantry deductions all pause for review.
Cleaner data over time | Unmatched ingredients go to an admin queue for resolution, not silent failure.#
Paste a URL → Aishi extracts title, ingredients, steps, and cultural notes → you review and confirm → recipe is saved and indexed for search.
Upload a receipt photo → OCR extracts line items → you review matches and quantities → confirmed items become pantry lots.
Chat opens the suggestion panel → RAG finds relevant recipes → results are ranked with transparent weights and short explanations → you pick one and move into cooking mode.
Cooking mode shows what will be deducted → you confirm → pantry lots are updated in FIFO order.
Aishi is both a RAG assistant and an AI agent:
LangGraph: workflows for chat, recipe extraction, receipt ingestion, and suggestions; tools route into deterministic flows; Postgres checkpoints so workflows survive restarts.
RAG: Recipe embeddings in ChromaDB; retrieval plus composite ranking (semantic + pantry + expiry); LLM-generated rationale grounded in scores.
Design choices that matter for real use:
Human-in-the-loop: Workflows interrupt before commits; idempotent resume prevents duplicate pantry writes.
- Cost-aware enrichment — Ingredient lookup hits the database first; the model is used only for unresolved items.
- Ethics by design — Allergen and food-safety risk is treated as high-impact; sensitive actions require explicit human approval.
- Home cooks who want one place to manage pantry, recipes, and meal ideas.
- Capstone / portfolio audiences interested in production-shaped agent design: HITL, auth, RAG, and multi-step workflows.
- Admins / curators who maintain the ingredient dictionary and resolve unmatched items from recipes, receipts, or manual pantry adds.
FastAPI · React · PostgreSQL · LangGraph · ChromaDB · OpenAI-compatible LLMs (chat, vision OCR, embeddings) · Docker / Render deployment
