Winning Hack@Brown with AdVantage
A 36-hour sprint with Tom Almog, a multi-agent ad system, and Best Multi-Agent Workflow on the Fetch.ai track.
This past weekend at Hack@Brown, my teammate Tom Almog and I won Best Multi-Agent Workflow on the Fetch.ai track with AdVantage: a full end-to-end agent system that helps small business owners research local competition, find gaps in the market, and generate entire ad campaigns from a single prompt.
View the full Devpost write-up →
Demo
Architecture
The pitch
Most small businesses cannot afford the $5K to $50K agencies charge for strategy and storyboards, so they skip real research and hope their ads work. AdVantage flips that: you describe your business in plain language, it runs location-based research (competitors, reviews, trends, SEO), then delivers three things in one flow. A research PDF with market gaps and keyword intel, a concept package (storyboard, script, budget, filming + hiring guide), and a ready-to-post viral clip for TikTok and Reels. Everything ships through ASI:One and saves to Google Drive.
What we built
A modular multi-agent pipeline where each stage is swappable:
- Marky (research): SerpAPI, Google/Yelp reviews, Google Trends + DataForSEO for seasonality, and a competitor map via Google Maps
- Script: Gemini parses your input; a trend analyzer and script writer turn research into an ad script
- Visuals: Vertex AI Imagen for pencil-sketch storyboards; FFmpeg + VEO 3, Lyria, and Google TTS for the final viral clip
- Delivery: ReportLab for the PDF, a Google Drive MCP agent for uploads, and the orchestrator dropping thumbnails + links in chat
Stack: Python, Fetch.ai Chat Protocol, Agentverse / ASI:One, SerpAPI, Vertex AI, Gemini, ReportLab, FFmpeg, Google Maps APIs, Google Drive MCP.
The 36 hours
We got paid $500 big ones to stay up 36 hours straight, shipping new ideas every couple hours and nearly crashing out at 4 a.m. A few things we hit and fixed: Imagen rate limits (added a 30s delay between calls), black video thumbnails (sample frames, pick the first non-black one), VEO/Lyria/TTS ordering, and robotic UX (rewrote copy in plain language with progress check-ins).
What I took away
Market research and creative generation work best as separate agent stages, and Fetch’s Chat Protocol made image previews and thumbnail-plus-links easy in ASI:One. Progress callbacks and graceful fallbacks kept the demo from dying on stage. Back-to-back hackathon wins with Tom have been a blast, and I am already looking forward to the next one.
Shoutouts
Huge thanks to Rajashekar V from Fetch.ai for the workshops, dev help, and merch. If you are building multi-agent systems, reach out. Always down to talk agents, ads, and sleep-deprived demo prep.