I built a Creative Director Agent for the Meta Global AI Developer Hackathon. It checks every product photo, builds ads with SAM 3.1, Muse Spark and Muse Image, and a human approves each one.
Code: https://t.co/onrqksZDqO
Live: https://t.co/lfmKJlOwLR
@theCryptoClock BitPOS lets the server set prices and the owner approve promotions. When a customer asks Muse Spark for a discount or an off-menu item, does it refuse on its own or pass the request to the owner?
@ChanochRosen InvoicePilot only pays the bank account already on file. When a supplier's real bank details change, how does it tell a genuine update from a fraud attempt without a person in the loop?
@imselmon When Matte reviews its own cut, what does it check against: your original sentence or the plan Muse Spark wrote from it? And is there a cap on how many re-edit passes it can run?
@RoketsSI Does Muse Image repaint the product itself in the 3 staged scenes, or only paint around the SAM 3.1 cutout? Label text on packaging is usually the first thing that drifts.
How the Creative Director Agent stays bounded: 3 review rounds and 4 image generations per run, then it stops. A critic model reviews each ad like an art director, and nothing is exported until a person clicks Approve. MIT licensed: https://t.co/onrqksZDqO
AI ad captions that only describe what the photo shows read as slop. A blue bottle on a table tells nobody why to buy. In the Creative Director Agent the model writes the headline only, plain code checks every line against the seller's facts, and a person approves every ad.
@avcs06 When the reviewer is the one judging the fixer's refusal, what stops it from always siding with its own earlier finding, and do you log how often a refusal is upheld?
@wellagore When a rider taps SOS with no signal on a mountain road, does GoBareng queue the alert and send it once signal returns, or does the 2 km alert only work while the phone is online?
@EdtechVara When GuardianCall hears a caller push for secrecy, like "don't tell anyone", does it interrupt the call itself or only prompt the family member to pause and phone someone they trust?
@Ming_LLM In a crowded dashboard like Meta Ads Manager, does DuckyHelper circle the exact field when I ask where to change the daily budget, or does it only describe where that field is?
@fatihu__m When the product picture has a printed label or logo, does SpreadFast keep those pixels exactly as uploaded in the finished ad, or does the model redraw the product?
@williamkast_ When brands get sued over misleading AI ads, is it usually the product looking different from what ships or a claim in the copy, and which of the two do you check by hand first?
I built an ad agent that is not allowed to invent facts. It checks every product photo, then SAM 3.1, Muse Spark and Muse Image build the ad, and a person approves before export. Made for the Meta Global AI Developer Hackathon. https://t.co/2dJF3T7cue
I built a Creative Director Agent for the Meta Global AI Developer Hackathon. It checks every product photo, builds ads with SAM 3.1, Muse Spark and Muse Image, and a human approves each one.
Code: https://t.co/onrqksZDqO
Live: https://t.co/lfmKJlOwLR
@nikolareljin When a label on an old part is worn or only half readable, does it guess the part number, ask for another photo, or mark the part unknown? A wrong read would send the whole build in the wrong direction.
@vinodvv How does reel2AIcart match what appears in a video to the right item in a Shopify catalog with pgvector: embed whole frames, or detect the product first? And how does it separate near identical colours or sizes?
@IvanRomanteev With approval bound to the exact SHA-256 plan hash, what happens when a timeout forces the planner to change one step? Does the human re-approve the whole plan, or can the changed step be approved on its own?
@CChang9909 On the leaking toilet request, how does Zukki settle on the likely fault and a price from Mario's own list when the customer video is unclear? Does Mario see a confidence level before his five minutes run out?
@iamwaqarjaved Thanks for the detail. Did Guard catch the one attack the unguarded agent missed? For the SAM misses, would a low confidence mask go to human review instead of a hard block?
@LittleCreatorFr How do you keep the images from the Muse Image agent consistent with the art direction in the GDD agent's document as the pipeline hands off between agents?