TrustLens

• 

  Architecture Diagram

Inspiration

The internet is full of information — and misinformation. From viral videos and social media claims to educational content, users are constantly exposed to unverified or misleading statements. Detecting and validating truth is a time-consuming and complex problem for everyday users.

TrustLens was born from this challenge — to empower the users to not only fight misinformation but also ensure the quality and credibility of the content they create or consume. TrustLens brings factual verification to everyone — educators, journalists, content creators, and everyday users.

What It Does

Media Processing: Extracts text from audio and video inputs using AWS Transcribe.

Claim Extraction: Identifies factual claims for verification from the text input or the transcripts.

Verification: Determines whether each claim can be validated using the model’s internal knowledge or requires external research, and then verifies the claims accordingly.

Research: Performs concurrent research using a thread pool to satisfy multiple queries in parallel, assisted by a Summarizer Agent that condenses tool outputs and prevents token overgrowth.

Response Generation: Generates a structured and clear output based on analyzed factual claims containing executive summary, list of analyzed claims, and suggestions .

How I built it

• Frontend: React + Tailwind CSS.

• Backend: Powered by AWS API Gateway (WebSocket API) and two Lambda functions:

  • Entrypoint Lambda — validates requests, generates presigned URLs for uploads, and invokes AgentCore Runtime.
  • Worker Lambda — triggered when AgentCore Runtime completes a task asynchronously and pushes results to Amazon SQS. This Lambda function then processes the verification output and sends the final response back to the client through the WebSocket Gateway

• Storage: Amazon S3 for secure media storage.

• AI Stack:

  • AWS Strands SDK for building specialized agents.
  • AWS Bedrock AgentCore for deploying and managing them.
  • Models: Claude Sonnet 4.5, Claude Sonnet 4, Claude Sonnet 3.7, and Claude Haiku 3.5 offered by AWS Bedrock.
  • Agents:
    • Orchestrator Agent: routes and coordinates verification tasks.
    • Media Transcriber Agent: handles transcription and preprocessing.
    • Claim Extractor Agent: identifies factual claims for verification from the text input or the transcripts.
    • Fact Reviewer Agent: determines whether each claim can be validated using the model’s internal knowledge or requires external research, and then verifies the claims accordingly.
    • Researcher assistant: performs concurrent research through a pool of researcher agents, assisted by a Summarizer Agent that condenses tool outputs to prevent token overuse.
    • Writer Agent: generates a structured response based on the input type and language.

• AWS Transcribe: Converts audio and video into timestamped text for the claim analysis and verification.

Challenges I Ran Into

• Excessive input token usage by the Researcher Agent — increased latency and cost, and sometimes exceeded model limits. Solved by adding a Summarizer Agent to condense tool outputs before passing them along.

• Inconsistent model behavior — improved through numerous iterations of prompt engineering, Guardrails, and structured outputs.

• Handling long-running verification tasks: Resolved by making AgentCore Runtime perform verification tasks asynchronously and push results to Amazon SQS, allowing Lambda Workers to finalize the process and send the results to the clients.

Accomplishments That I am Proud Of

• Developed an intelligent multi-agent system that autonomously coordinates reasoning, research, and fact verification.

• Achieved >70% reduction in researcher agent input tokens — leading to a >70% reduction in model cost through intelligent summarization and context management. View case study →

• Implemented parallelized research, drastically improving verification speed.

• Implemented a scalable, event-driven backend architecture that cleanly separates synchronous request handling from asynchronous agent execution — eliminating timeout issues entirely.

What I Learned

This was my first experience building an AI agent, and it taught me a lot about multi-agent systems and orchestration. I gained hands-on experience with AWS Strands and Bedrock AgentCore.

What’s Next

• Integrating additional data source connectors for scientific and journalistic verification.

• Implement user authentication and personalization to deliver a more tailored experience.

Built With

• agentcore
• calude
• lambda
• python
• s3
• strands
• transcribe