Project Pulse

Pulse System Architecture

Inspiration

When an unscripted cultural moment happens. say superbowl final touchdown or a Google I/O event, consumer attention spikes to its absolute peak. But the traditional enterprise marketing apparatus is slow. It takes an agency hours to secure legal approvals, draft copy, and render media. By the time they launch, the cultural window is dead. You can't just be fast anymore; you have to be immediate. We were inspired to bridge this gap and build a system that turns cultural velocity into zero-latency revenue, completely autonomously.

What it does

Pulse is a real-time, event-driven cultural marketing engine. When a viral moment occurs, marketers inject a simple text trigger into the Pulse dashboard. From there, Pulse takes over entirely:

Enterprise Grounding: It queries our BigQuery data warehouse for our brand’s highest-performing historical ads to use as RAG context.
Multimodal Generation: The primary AI agent (The Creator) generates an entire "Infinite Canvas" campaign matrix - including localized ad copy, highly specific image prompts, and voiceover scripts.
Autonomous Compliance: A secondary agent (The Brand Guardian) intercepts the draft, auditing it for IP and copyright violations and autonomously rewrites the prompts to ensure legal safety. And finally generates a review score.
Asset Creation: Pulse triggers Vertex AI Imagen and GCP Text-to-Speech to generate the actual media, automatically embedding DeepMind SynthID watermarks for Responsible AI compliance.
Execution: It formats the final assets into a Meta Graph API-ready JSON payload, streamed in real-time to our Next.js dashboard for one-click publishing.

How we built it

We built Pulse entirely on Google Cloud Platform to ensure enterprise-grade scalability: Frontend: A Next.js (App Router) dashboard hosted on Cloud Run. Eventing: Triggers are pushed via Cloud Pub/Sub, allowing for a decoupled, scalable architecture. Orchestration: Node.js microservices running on Cloud Functions (Gen 2) act as the brain of the system. The AI Core: We utilized the Gemini Pro Multimodal API for our multi-agent orchestration, heavily relying on its Interleaved Output capabilities to generate text and JSON payloads simultaneously. Media & Data: We used Vertex AI Imagen for visuals, GCP Text-to-Speech for audio, Google BigQuery for our RAG database, and Firestore to stream the generated campaigns back to the UI in real-time.

Challenges we ran into

Our biggest hurdle was getting the Gemini model to output a strictly formatted, complex JSON matrix that could reliably trigger the downstream media APIs without breaking the frontend parsing. Early iterations hallucinated keys or added conversational markdown. We solved this by implementing strict few-shot prompting and system instructions, eventually achieving a 100% success rate on JSON formatting. Additionally, we struggled with "generic AI voice." AI-generated ad copy often sounds robotic. We solved this by implementing our BigQuery RAG pipeline, forcing Gemini to analyze past high-ROAS enterprise campaigns before generating a single word, ensuring the brand voice remained authentic.

Accomplishments that we're proud of

We are incredibly proud of The Brand Guardian. In enterprise software, speed is dangerous if it isn't legally compliant. Building a secondary Gemini agent that actively intercepts, critiques, and rewrites the primary agent's prompts to prevent copyright lawsuits feels like a massive leap forward in agentic workflows. Emitting those structured safety interventions directly to GCP Logs Explorer makes the platform feel truly production-ready. We are also proud of our strict adherence to Responsible AI.

What we learned

We learned the sheer power of the Google Cloud ecosystem when the pieces are connected properly. Realizing that we could seamlessly pass data from a Pub/Sub topic, through a Gemini-powered Cloud Function, into a Vertex AI media generator, and stream it live to a frontend via Firestore - all within milliseconds. We learned that the future of AI isn't single-prompt chatbots; it is multi-agent, serverless orchestration.

What's next for Project Pulse

We are currently building the "Hawk-Eye" Vision Integration. Right now, Pulse relies on manual text triggers. Next, we will integrate the Gemini Vision API to autonomously ingest live YouTube .mp4 streams. Instead of waiting for a human to type a trigger, Pulse will "watch" a competitor's live keynote stream. The exact second a competing product price is revealed on screen, Pulse will visually extract the data and autonomously deploy a counter-campaign before the keynote even ends.