How AI Cracked the Code on Viral Restaurant Content

An Austin-based social media agency manages content strategy and performance for a portfolio of restaurant clients across Texas. Their competitive edge rests on a proprietary six-element virality framework — covering Hook, Topic, Value, Format, Call to Action, and Editing — developed through years of hands-on restaurant marketing. As short-form video became the dominant content format and their client roster expanded, the agency reached an inflection point: the institutional knowledge that made their work exceptional was locked inside individual team members, and the data needed to validate and scale it was buried in disconnected platform reports that took hours to compile. The organization recognized an opportunity to transform their creative methodology into a durable, data-driven competitive advantage — one that could grow with the business rather than depend on any single person.

The project aimed to deliver a proof-of-concept media analytics platform with two core capabilities: a centralized data lake aggregating social performance data across all client accounts, and an AI-powered video analysis engine that could evaluate content against the agency's six virality attributes. Success meant client managers could replace manual reporting with real-time dashboards and receive AI-generated recommendations grounded in historical performance patterns — turning creative intuition into a repeatable, scalable system.

The engagement required solving two fundamentally different data problems in parallel. On the analytics side, the agency's existing social media management platform capped cross-account reporting at 12 profiles and offered no agency-wide view — meaning every performance review required manual data exports, spreadsheet compilation, and time-consuming reconciliation across accounts. Building a data lake capable of ingesting, normalizing, and persisting this data at scale — while keeping dashboards accurate and metrics like reach, views, and follower growth correctly calculated — demanded careful data modeling and rigorous quality validation throughout.

‍

On the AI side, the team faced the harder challenge of operationalizing a subjective creative framework. The six virality attributes are nuanced and context-dependent: distinguishing a strong visual hook from a weak one, or evaluating editing pacing, requires the kind of domain judgment that resists simple rule-based automation. Processing short-form video ranging from seven to sixty seconds through multiple analytical lenses — frame extraction, object detection, audio separation, transcription, and on-screen text recognition — and synthesizing those signals into actionable attribute scores demanded a multi-agent architecture with precise prompt engineering at each stage.

Protagona built a cloud-native data lake on Amazon S3 — using AWS Lambda for ingestion and transformation — to consolidate social performance data across the agency's full client portfolio into a single, continuously refreshed source of truth. The architecture retains historical data independently of upstream platform changes, solving a long-standing problem of lost performance history when accounts or integrations changed. Amazon QuickSight served as the dashboarding layer, giving client managers real-time visibility into reach, engagement, follower growth, and post-level performance — no engineering support required.

‍

For video intelligence, the team built an AI processing pipeline on Amazon Bedrock to score short-form restaurant videos against the agency's six virality attributes. Each video passed through specialized functions covering visual content, on-screen text, audio, and transcription, before a synthesis layer converted those signals into structured attribute scores. A statistical analysis function then correlated creative characteristics with actual reach outcomes, surfacing which hook styles, formats, and calls-to-action historically drove the highest performance by client type.

‍

To enable conversational querying, the team configured Amazon Q to draw from curated dataset topics and dashboard visualizations — ensuring natural-language queries returned answers grounded in the data lake. The result: a unified intelligence platform where dashboards, AI scoring, and conversational querying all operate from the same trusted data foundation.

‍