# Brieblue

- **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:** [Edward Prado](https://cerebralvalley.ai/u/Brieblue)
- **GitHub:** https://github.com/Briebluedotcom/brieblue-multilingual-ad-factory
- **Demo video:** https://youtu.be/HUsung_300g
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2106

Project Description — Brieblue Multilingual Ad Factory
INTRO:
"Hello my name is Ed Prado from Brieblue. We are not an AI company; we are an art and textile importer. We are addressing a very real problem affecting our world. Our world may consist of traditional Hmong artists supplying us from Chiang Mai, Thailand, a Filipina customer shopping in San Antonio, Texas, or my own family celebrating our Chinese heritage. Our world literally is the world."
BRIDGE:
"With the advent of AI, small businesses like mine are increasingly able to operate in a global environment. When we first started out in 2014, artists still relied on galleries to market their works. Facebook and Instagram made it possible for micro businesses to reach audiences never before possible. This creates some unique problems."
PROBLEM:
"Here in Texas, we participate in Asian cultural festivals. Attendees may be Russian, Thai, or Korean. Our suppliers in Thailand may be trying to reach buyers in Vietnam or Indonesia. Or maybe trying to break into the American or Chinese markets. Unfortunately for all of us, the default is to market in English. We've all seen the poorly translated advertisements that make advertising agencies seem like a joke. Everyone speaks a little English. The problem is it's nobody's native language. We set out to change the way small businesses market in today's global environment."
ORIGIN
"Many years ago I noticed that when traveling in Thailand, my feed becomes dominated by local content. Bright, happy, unique. When I return home, my feed gradually returns to normal. Safe, boring, corporate. This pattern is clearly visible in the two devices that make our business possible. An iPhone 12 with service in the United States, and a Vivo V60 registered in Thailand. One bright and cheery, the other more utilitarian. Both essential to our daily operations. These differences have me longing to return to the comfort of my second home. We quickly realized that many of our customers felt the same way. A longing for home and a desire to connect with people of similar beliefs unites our community. There is currently no effective way to include multiple linguistic and cultural elements into a singular advertising campaign. Part of our motivation to post on social media has always been to bring the joy we experience from art to others and to try to include all cultures as we build for our world."
IDEA
"Take a still English-language event flyer and build thousands of personalized versions for each of the communities participating in our events."
HYPOTHESIS
"To apply Meta's global strategy adjusted for small business."
CRITERIA
"We are mobile only."
PRACTICE
"We borrowed the idea from Meta to hire country-specific managers to handle dedicated tasks. Except that for our purposes we used AI assistants. When our 'mail department' receives an event flyer by email, it automatically animates elements of the flyer to a selected tempo. This master copy is then sent to participating country managers for adaptation. Country managers then select from 8 possible local musical accompaniments whose licenses range from organic use only to full ownership. These 8 versions can then be additionally modified for visually impaired or non-reading users."
Country assistants run on Muse Spark 1.3. Flyer subject isolation uses SAM 3.1 via the Meta Model API.
DIWALI BACKSTORY
"Part of the inspiration for this project comes from a prior campaign for the Diwali San Antonio celebration. We took the original event flyer and produced 10 separate advertisements for each Indian language supported by Meta. For this campaign, produced on October 6, we used licensed tracks automatically procured through YouTube. By the start of the hack, this access had been removed. For the hack, we wanted to expand this concept by creating individual agents for each country represented at the AANHPI Night Market in Austin, Texas. We also wanted to replace the loss of agentic access to the music library as well as address the licensing issues involved with creating a paid ad campaign."
MUSIC PATHS
"In order to accomplish this, we examined various methods of obtaining music for our ad factory. With loss of agentic access, one option was to allow country agents to select songs that could be manually selected on various platforms. Users could also upload tracks from their personal libraries, but these options do not solve the licensing issue. An option to avoid licensing is to use recordings made pre-1923. These can then be remixed to produce more modern sounding tracks to fit current audiences. However, finding relevant media can prove difficult. For meta specific campaigns, Meta Sound Collection can be used. However, no agentic access for this could be found either, and search selection can be limited. Each country presented with multiple options, building in redundancy. Overall the most scalable option which retains ownership was to generate our own tracks."
Meta Sound Collection requires manual browsing — no API access was found.
MUSIC GENERATION
"In order to accomplish this, we deployed two separate music generation tools, a simple synth usable by Spark, and full ACE-Step generated on Modal with GPU. These tools received creative instruction from country agents in combination with the main agent responsible for video editing."
The synth is orchestrated by Muse Spark 1.3. ACE-Step (third-party, Apache 2.0) runs on Modal GPUs.
CAPTIONS + SPOKEN VERSE
"Once the track is selected, country-specific agents can then add live captions and spoken verse in their native languages. This is potentially the most valuable tool, reaching visually impaired and non-reading users in their native languages. If the selected track contains vocals, Demucs is used to allow for voiceover."
Demucs is Meta's open-source stem separation. Narration QA uses Muse Voice Transcribe via the Meta Model API.
PROOFREADING
"Once text and voice are added, the draft is sent to our proofreading assistant to confirm completion prior to submitting to the user for final approval."
Hard cultural and licensing calls escalate to Muse Wonder, the max-reasoning tier.
AGENT ARCHITECTURE
"For the Asian festival campaign undertaken for this event, we currently have six active language agents with 24 open positions to be filled as needed. This can be expanded to the 71 languages supported by Ads Manager and eventually beyond to the 200 languages included in NLLB — Meta's No Language Left Behind translation model. Each of these agents is specifically designed to provide linguistic and cultural insights to their assigned region. They can provide active text, spoken verse, and captions in their assigned language, as well as make specific cultural recommendations for country specific marketing campaigns.
Another key feature of this factory is the agent's ability to add live text and voiceover. This concept was an essential element to the project due to my personal experience. Having visually impaired family members, I understand the difficulty faced by the visually impaired in an increasingly technological environment. Also, in some lucrative markets as much as twenty percent of the population does not read. These agents can conduct direct marketing campaigns through WhatsApp and email including visually impaired and non-reading users through direct verbal message. They can also reach users in China, Myanmar, and Russia without Meta access through Microsoft Outlook campaigns. And the email channel works both ways — in countries where Muse itself is not available, a user can send a flyer by email and receive the finished localized ad back the same way. The back-and-forth email portal is the workaround for the availability gap, not just an outgoing broadcast channel. This tool was tested using four available Meta API models. Muse Spark 1.3 orchestrated the build. Reasoning was tested on Muse Spark Open using Baseten credits. Glimmer was used to test factory models using serverless SFT on Fireworks — 19 pairs on Oct 30, then 25 pairs on Oct 31 after adding Ed's verdicts as training labels. And Muse Wonder was used for difficult calls. In all, a 25 pair training set with a 3,200-pair vision was conducted for verification. The tuned model answered the holdout test in 805 tokens versus the base model's 6,579 — more concise, having learned to get to the point."
PLATFORM AGENTS
"Once a final copy is generated, it is handed off to our proofreading agent who confirms message and visual continuity, before being submitted to the user for final approval. An independent agent designed for human-like verification is a concept we built into our Diwali campaign. However, for Asian Festival this design was further distilled for specific proofreading tasks. If approved, this final copy can then be sent to any number of agents who can specifically tailor the advertisement for particular platforms. For this campaign, we designed an agent specifically with the parameters of the X platform in mind for the viral post category. Additional positions remain unfilled for 11 other platforms like TikTok and LINE."
SCALE + FUTURE
"Have been asking for this for years. With the release of Muse, I just did it myself."
"Using all potential pathways, it is possible to turn one static image into almost 77,000 usable advertisements. By designing each agent with specific tasks, we can create an efficient architecture to run a global advertising campaign from a mobile device. For future campaigns, we hope to expand the languages supported. One limitation we face, operating in art shows around the world is access to computing power. Our current operation consists of an iPhone 12 and a Vivo phone. Further implementation would include full roll out of a three-tiered system consisting of Wonder operating through the Meta API, and a home server utilizing a Muse open-source, all accessed through a mobile device on Spark 1.3. The home server would replace Spark Open on Baseten and Glimmer. Each tier responsible for specific tasks. In our opinion, an ideal international test bed for this ad factory would be Singapore. Again following Meta's example of basing their Asia operations in Singapore and using this as a first market to access products. This provides a unique environment with multiple native languages. Future duties for Country specific agents would include in country advisement and market monitoring"
The scale math: 76,800 = 8 paths × 100 languages × general/read × 12 platforms × 4 models.
CONCLUSION
"When we started Brieblue in 2014, the intention was to build a cultural bridge between Texas and Asia. Allowing independent artists to reach new markets in new locations without ever having to change physical locations. But the reality was that our audience was so fragmented that there was no easy way to market to our community. It has taken us years of doing in-person events to find the local customers that have been among us all along just waiting to be found. We're have been asking for tools like this in the Meta business leaders forum for some time. With the release of Muse we are able to produce them ourselves. We believe that artificial intelligence together with our application can allow local users to reach communities never before available. It can also enable micro businesses to produce a globally inclusive advertising campaign from a mobile device. The impact of this? Probably putting me out of business when artists in Bangkok can target audiences in San Antonio. But that was always the point."

Contributions disclosure (for submission)
Built during the hackathon (Oct 26–Nov 4, 2026): Asian Festival expansion to all Asian languages, 6 country-specific agents + 24 open positions, 8 music paths including owned tracks, all synth beds (Oct 28), Mail Dept / email portal, Proofreading Dept, platform-specific agents, and all Meta model integrations (Spark 1.3, Spark Open, Glimmer, Wonder, SAM 3.1, Voice Transcribe). Pre-existing (disclosed, shown for context only): original Indian songs from YouTube (closed access); the cut-chart concept; the Diwali factory as an India-only design. The demo video shows only hackathon-built work; the still flyer appears as input only.

## More from Meta Global AI Developer Hackathon

- [Kepler](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2103)
- [Devisi Coder](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2104)
- [PROPERTY OS](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2105)
- [NULLPTR](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2107)
- [Axis](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2108)
- [strollcall](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2109)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2106. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt