Build a Hyper-Personalized Voice Assistant Agent
This challenge tasks developers with creating an advanced, real-time voice assistant designed for hyper-personalization and proactive device or application management. The solution must leverage Mastra AI's agentic workflows and robust memory capabilities to maintain long-term user preferences and context. Claude Opus 4.6 will be integrated for its nuanced conversational understanding and empathetic response generation, while ElevenLabs will provide natural, low-latency speech synthesis and recognition for a seamless voice user experience. The assistant should dynamically adapt its responses and actions based on the user's historical interactions and real-time device state. Participants will focus on designing reactive agent workflows in Mastra AI, implementing bidirectional real-time speech interaction, and developing custom tools for device integration. The ultimate goal is to deliver a highly intuitive and personalized voice interface that anticipates user needs and acts intelligently, enhancing the user's interaction with their digital environment.
What you are building
The core problem, expected build, and operating context for this challenge.
Create a Mastra AI agent for real-time, hyper-personalized voice assistance, integrating Claude Opus 4.6 for natural language and ElevenLabs for speech.
How work is evaluated
Evaluate the voice assistant's ability to accurately understand complex voice commands, leverage historical context for personalization, and execute tasks reliably in real-time.
Challenge datasets
Public datasets, plus the private datasets you have access to.
How submissions are scored
These dimensions define what the evaluator checks and which criteria separate a passable run from a strong one.
task_completion_accuracy
Checks if all requested tasks were completed correctly and without errors based on the user's voice commands.
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
memory_recall_effectiveness
Verifies if personalized information from long-term memory was correctly accessed and utilized in responses or actions.
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
real_time_responsiveness
Assesses if responses were generated and spoken within acceptable latency for a natural voice interaction (e.g., < 1 second).
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
personalization_score
Degree to which the assistant adapted its behavior and responses based on user preferences and history. • target: 0.85 • range: 0.6-1
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
conversational_turn_accuracy
Percentage of conversational turns correctly understood, interpreted, and responded to by the assistant. • target: 0.92 • range: 0.8-1
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
average_response_latency_ms
Average time taken (in milliseconds) from user speaking to assistant starting to respond. • target: 300 • range: 50-1000
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
What you should walk away with
Master building reactive agent workflows with Mastra AI, utilizing its state management and tool orchestration for dynamic task execution.
Implement real-time bidirectional speech interaction using ElevenLabs Text-to-Speech and Speech-to-Text APIs for a seamless voice user experience.
Design and integrate a long-term memory system within Mastra AI to capture and recall user preferences, historical interactions, and device context for hyper-personalization.
Leverage Claude Opus 4.6's advanced conversational capabilities for nuanced natural language understanding, intent recognition, and empathetic response generation.
Develop custom tools for the Mastra AI agent to interact with simulated device settings (e.g., calendar, reminders, app control) and external APIs.
Integrate the voice assistant with Ellipsis to provide a rich conversational UI alongside the voice interface, enhancing accessibility and interaction modes.
Implement robust error handling and conversational recovery strategies within the Mastra AI agent to maintain fluidity during complex interactions or misunderstandings.
Reference links and supporting material
Sample data for 2 tasks
Challenge setup
Evaluate the voice assistant's ability to accurately understand complex voice commands, leverage historical context for personalization, and execute tasks reliably in real-time.
Challenge input
JSON object with 'user_voice_commands' (array of strings) and 'initial_memory_state' (object representing stored preferences).
Agent execution
No agent tool is listed for this challenge.
Evaluated output
JSON object with 'actions_taken' (array of strings) and 'personalization_score' (float).
- Checks if all requested tasks were completed correctly and without errors based on the user's voice commands.
- Verifies if personalized information from long-term memory was correctly accessed and utilized in responses...
- Assesses if responses were generated and spoken within acceptable latency for a natural voice interaction (...
- Personalization Score target: 0.85
- 2 public reference cases
- JavaScript execution harness
Summary shows the first task, up to three tools, and up to three evidence items. Open the technical recipe to view all configured items.
View technical recipe
Configured tools
No tool records are attached.
Evaluation contract
- task_completion_accuracy · Weight 1
- memory_recall_effectiveness · Weight 1
- real_time_responsiveness · Weight 1
- personalization_score · Weight 1
- conversational_turn_accuracy · Weight 1
- average_response_latency_ms · Weight 1
Metrics
- personalization_score · Target 0.85
- conversational_turn_accuracy · Target 0.92
- average_response_latency_ms · Target 300
Reference cases
- 2 public reference cases
Workflow tasks
- Task 1
Input: JSON object with 'user_voice_commands' (array of strings) and 'initial_memory_state' (object representing stored preferences).
Output: JSON object with 'actions_taken' (array of strings) and 'personalization_score' (float).
Execution harness: JavaScript
- Task 2
Input: JSON object with 'user_voice_commands' (array of strings) and 'device_state_log' (array of initial device actions).
Output: JSON object with 'final_state' (string) and 'conversational_fluency_score' (float).
Execution harness: JavaScript
Recipe freeze marker
No recipe freeze time is recorded.