ActionLens

Inspiration

Most AI assistants stop at answering questions or summarizing documents. Real-world documents often contain deadlines, requirements, risks, and tasks that still require manual work.

We built ActionLens to turn that process into an agentic workflow. Instead of simply summarizing a document, ActionLens understands the document together with the user's intent, chooses the most relevant workflow, identifies important actions and risks, and persists actionable tasks for later use.

What it does

ActionLens is a document-aware and intent-aware AI agent built for the Taskmaster track.

It can:

• Accept and process PDF documents.
• Understand the document context together with the user's intent.

• Dynamically route requests to specialized workflows:

  • Student Opportunity
  • Research & Technical
  • Privacy & Terms
  • General

• Extract important findings, missing information, deadlines, and risks.

• Create actionable tasks in Google Cloud Firestore.

• Read back the stored data to verify that the action items were successfully persisted.

This allows ActionLens to move beyond a traditional document chatbot toward an agent that can actually organize the next steps from complex documents.

How we built it

ActionLens uses Google Agent Development Kit (ADK) as the agent framework, with Gemini running through Vertex AI for reasoning and workflow decisions.

The application is containerized and deployed on Google Cloud Run, while Google Cloud Firestore provides persistent storage for generated action items.

Technology Stack

• AI: Gemini 3.6 Flash via Vertex AI
• Agent Framework: Google Agent Development Kit (ADK)
• Cloud Infrastructure: Google Cloud Run
• Database: Google Cloud Firestore
• Document Processing: Python + pypdf
• Language: Python

The architecture separates document processing, agent orchestration, specialized workflows, and persistent action storage so that additional document types and workflows can be added in the future.

Challenges

One of the main challenges was making the agent adapt its behavior based on both the document and the user's intent instead of forcing every document through the same workflow.

We also wanted generated actions to be more trustworthy than simple model output. For this reason, ActionLens writes action items to Firestore and performs a read-back verification to confirm that the data was actually persisted.

Testing also included ambiguous documents, documents with very little information, conflicting data such as different deadlines, and multilingual/Hinglish content.

Accomplishments

We are proud to have built an end-to-end agentic workflow that:

• Dynamically routes documents to specialized workflows.
• Identifies actionable information, missing details, and risks.
• Persists generated action items to Google Cloud Firestore.
• Verifies persistence through read-back confirmation.
• Runs as a containerized application on Google Cloud Run.
• Uses Google's ADK and Gemini through Vertex AI as core parts of the system.

The result is ActionLens — an agent that doesn't just tell you what a document says, but helps turn it into what you should do next.

Built With

• agent
• docker
• document
• documentprocessing
• firestore
• gemini
• gemini-api
• generative
• google-adk
• google-cloud
• google-cloud-firestore
• google-cloud-run
• pypdf
• python
• run
• vertex-ai