# Synteron - **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon) - **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST) - **Where:** Online - **Team:** [Sabhyata Chaudhari](https://cerebralvalley.ai/u/SabhyataC) - **GitHub:** https://github.com/SabhyataC/MarketBrain - **Demo video:** https://youtu.be/_NyjvZmg9EA - **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery - **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2193 The problem A small shop owner usually has no marketing team. Listing products takes hours of typing. Writing ads, search text and landing pages takes skills and time. When they try AI tools, they have good reasons not to trust them, because an AI can invent claims, overspend, or publish something wrong. So most small shops get no marketing support at all. What Marketing Brain does Marketing Brain turns one video of a shelf into a reviewed catalog, ads, a four-picture ad series, a landing page, and search listings for the owner's shop. The central idea is simple: the AI proposes, plain code checks, and the owner approves. Nothing goes live, nothing is sent to a shop or to Meta, and no money is spent until the owner says yes. Every action is logged, and every change to a shop can be rolled back. How it works 1. Camera to catalog. The owner films a shelf with price tags. The app takes one frame per second, and Muse Spark (vision) reads every readable tag. The same product appears in many frames, so readings are merged by brand, name and price. A product read the same way in two or more frames is trusted. A product seen once, or read two different ways, is flagged for the owner. The owner then corrects the draft and approves it. 2. Planner. Muse Spark picks three products to promote, using the discount and the stock, and gives a reason for each. 3. Ad engine. Muse Spark writes three ad lines. Muse Image edits the product's own photo into an ad image, so the product in the ad is the product on the shelf. Rule checks block medical claims, guarantees, "best", and fake offers. A model check flags physical claims that are not in the product's facts, and the owner must confirm those. Budget caps are enforced in backend code. 4. Ad series. Four ads from one photo, with one shared style so they look like one campaign. The app, not the image model, draws the headline, price and button, so the text is always correct. A colour-consistency score is shown. One picture can also be translated into Hindi, Marathi, Gujarati, Tamil, Bengali or Telugu. 5. Landing page. The model writes only the words. The app's own template builds the page. Then a real browser opens it at phone and desktop size, takes screenshots, and measures seven things: no errors, no sideways scroll, the photo is visible, the price and button are visible without scrolling, readable contrast, a tappable button, and no overlap. A failed check is shown in red and never hidden. 6. SEO engine. The model proposes a title, description, heading and image text with a reason for each. Code checks length limits, keyword stuffing, prices and risky claims, and builds product schema from the real catalog. The owner sees a before and after diff, approves, applies, and can roll back. 7. Send to a shop. Products go out as drafts by CSV (Shopify, WooCommerce), a signed webhook (the owner's own site), or the Shopify and WooCommerce APIs. 8. Kill switch. One button pauses every approval, send, apply and model call. Where Muse is used Muse Spark (muse-spark-1.1) does the vision reading, planning, copy, translation, claim checks and SEO text. Muse Image (muse-image-1.0) edits product photos into ad and landing page images. All calls go through one logged client, and everything else (prices, budgets, rules, layout, measuring) is plain, testable code. Results and honesty On a 14-product test shelf, the app found all 14 products. Price was correct for all of them, and 13 of 14 had name, brand and price all correct. The one error was a misread name, and the app flagged it for review. Building the draft took 85 seconds, against about 28 minutes to list the same products by hand (an assumed figure, not measured). This is a single run on one clean video, so it is an upper estimate. The signed webhook was tested against a real Django project built from a real store's models. Shopify and WooCommerce were tested only against simulated shops, and I say so openly. Nothing is sent to Meta: approved ads are kept as paused drafts with a pack to upload by hand. What's next Whole-store planning, competitor-ad analysis, demand forecasting, translation of all four series pictures, a direct Search Console connection, and real-shop runs for Shopify and WooCommerce. Built with: Python, FastAPI, SQLite, ffmpeg, Pillow, Playwright (Chromium), the OpenAI SDK against the Meta Model API, and a single-page HTML interface. ## More from Meta Global AI Developer Hackathon - [Action Shift by Never86’d](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2190) - [Jarvis Team](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2191) - [SettleProof](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2192) - [Muse Yoga](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2194) - [Overcomplex](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2195) - [Mockagent](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2196) --- Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2193. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt