Example - Speech Coach (SPA)

 

 

Concept

Suppose a student team wants to create a speech coach app (SPA) where small videos can be uploaded and analyzed giving both a critique, improvement and practice options.

 

 

React SPA ↓ Node/Express Backend ↓ Upload short video ↓ Extract / access: • video frames • audio track ↓ Gemini direct multimodal analysis ↓ Structured observations ↓ Google ADK Agent Architecure          ↓         Coaching Critique / Improvement Suggestions / Practice Options ↓ Firestore

 

 

 

 

 

 

Agent Architecture:

 

The architecture is a hierarchical Google ADK multi-agent system built around an Orchestrator Agent. The application first sends each relevant section of the uploaded speech video to Gemini for direct multimodal analysis, producing structured observations about Speech Content, Vocal Delivery, and Body Language. Those Gemini analysis results become the input to the ADK workflow.

The Orchestrator Agent manages three specialized agents in sequence. The Critique Agent interprets the Gemini observations and presents clear, section-by-section critiques for Content, Vocal Delivery, and Body Language. The Improvement Agent consumes those critiques and converts them into specific, actionable recommendations for each category and video section. Finally, the Practice Agent uses both the critiques and improvement recommendations to create targeted exercises—for example, rewriting an opening, practicing pacing and pauses, or rehearsing eye contact and gestures. The Orchestrator combines the three outputs into the final coaching experience shown in the React App and can persist the analysis, recommendations, and exercises in Firestore for later comparison and progress tracking.

 

 

 

 

 

NOTE: The initial Gemini calls perform direct multimodal analysis of the video/audio. The ADK agents subsequently use Gemini as their reasoning model while operating on those analysis results and invoking any tools provided to the agents.The initial Gemini calls perform direct multimodal analysis of the video/audio. The ADK agents subsequently use Gemini as their reasoning model while operating on those analysis results and invoking any tools provided to the agents.

NOTE: This is one example of an appropriate agent architecture. Your Project 2 agent architecture should be designed specifically for the purpose of your application and does not need to follow this architecture.