Bano's Cookbook — preserving a family recipe archive

Live at banoscookbook.com

The story

A family cookbook started by hand in 1981. Four decades later, the paper is fading and the handwriting is getting harder to read. This is the project that gets it out of the notebook: 13 of my grandmother's recipes, transcribed and published — Bori Samosa, Roghan Josh, and the rest — each one attributed to her, each with generated imagery.

What is built

The visible part is a recipe site. The interesting part is the pipeline behind it:

• Handwriting extraction — Gemini flows that read a scanned page and return a structured recipe: ingredients, quantities, steps
• An eight-step video workflow — driven by an XState state machine, turning a written recipe into a scripted, narrated cooking video
• A containerised encoder — FFmpeg does not belong in a serverless function, so the video combiner lives in its own Cloud Run container
• Instagram integration that genuinely ran — real comments synced back from the Graph API

16 Gemini flows, a 5,900-line server actions module, and one very stubborn video encoder.

Honest status

Thirteen recipes were transcribed in October 2025, and the archive has been paused since — the 1981 book is not fully ingested. Every recipe in the database was entered by me or seeded as demo data; there is no external userbase and no paying customers. The video pipeline runs end to end in the app, but nothing it produced has been published to a recipe yet. TikTok publishing is scaffolded, not connected.

It is a single-tenant application with multi-tenant plumbing.

Why it matters for client work

Digitisation projects are won or lost on workflow design, not on the model you pick. This one covers the whole path — messy analogue source, structured extraction, human review, publishing — and it produced two genuinely transferable lessons that cost real days to learn: what an ordered Firestore query silently drops, and why you should store which code path produced a result.