Skip to main content
Back to challenges
Challenge

Multimodal Flood Resilience Agent with Google ADK and Llama 3.3 70B

Inspired by the recent devastating floods in West Virginia and the failure of unaddressed prevention plans, this challenge tasks developers with building a sophisticated Flood Risk Assessment Agent. Using the Google Agent Development Kit (ADK), you will create an agent that can process multimodal inputs: geospatial satellite imagery of property footprints and historical flood prevention documents. The agent must evaluate the 'resilience gap' between proposed prevention plans and current physical infrastructure. You will utilize Llama 3.3 70B (running on Vertex AI) to parse and summarize complex legislative and engineering reports, while the ADK manages the agentic loop and tool integration for fetching real-time weather alerts and elevation data. The core of the challenge is the implementation of a reasoning chain that can identify if a property's current mitigation measures (visualized via computer vision) align with the state-level prevention strategies. This involves high-stakes decision-making for insurers to adjust underwriting premiums based on 'active resilience' rather than just static FEMA zones.

Machine LearningHosted by Vera
Sign in free to start
Challenge brief

What you are building

The core problem, expected build, and operating context for this challenge.

Develop an automated flood risk scoring system that combines multimodal satellite analysis with Llama 3.3 70B text analysis using the Google ADK.

Delivery guide

How work is evaluated

Evaluation

Evaluation focuses on the accuracy of the resilience score and the agent's ability to cross-reference text with visual data.

Datasets

Challenge datasets

Public datasets, plus the private datasets you have access to.

Evaluation rubric

How submissions are scored

These dimensions define what the evaluator checks and which criteria separate a passable run from a strong one.

Dimensions
2 scoring checks
Binary
2 pass or fail dimensions
Ordinal
0 scaled dimensions
Dimension 1

Logic Consistency

Check if evidence mentions specific clauses from input text

Binary check

This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.

Dimension 2

Precision at 0.5

Accuracy of detecting non-compliant infrastructure • target: 0.85 • range: 0-1

Binary check

This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.

Learning goals

What you should walk away with

  • Orchestrate multimodal AI agents using Google ADK and Gemini models

  • Integrate Llama 3.3 70B for large-scale document parsing within a Vertex AI ecosystem

  • Implement geospatial risk scoring using raster and vector data overlays

  • Design an explainable insurance underwriting decision support system

Resources and assets

Reference links and supporting material

Dataset notes

Synthetic dataset containing high-resolution property images with ground-level vs. elevated HVAC systems, paired with semi-structured PDF reports of state flood prevention plans.

Challenge setup

Evaluation focuses on the accuracy of the resilience score and the agent's ability to cross-reference text with visual data.

Challenge input

JSON containing image_url and pdf_text

Google ADK

Used for managing agentic state and multimodal tool use.

Llama 3.3 70B

Provides deep understanding of legislative and engineering text.

Vertex AI

Host for model endpoints and compute.

Evaluated output

JSON with 'score' (0-1), 'violations' (list), and 'evidence' (string)

Evaluation checks
  • Check if evidence mentions specific clauses from input text
  • Accuracy of detecting non-compliant infrastructure • target: 0.85 • range: 0-1
Configured evaluation evidence
  • Benchmark: InsuranceML-Resilience-v1
  • Precision At 0.5 target: 0.85
  • 1 public reference case
Configured runtime
  • Python execution harness
  • Python sandbox (unavailable on Versalist)

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

Environment
  • Modal · Optional
Action Space
  • Google · Required
  • Llama 3.3 70B · Optional
Policy Serving
  • Modal · Optional

Evaluation contract

  • Logic Consistency · Weight 1
  • Precision at 0.5 · Weight 1

Metrics

  • Precision at 0.5 · Target 0.85

Benchmarks

  • InsuranceML-Resilience-v1
  • security

Reference cases

  • 1 public reference case

Workflow tasks

  1. Task 1

    Input: JSON containing image_url and pdf_text

    Output: JSON with 'score' (0-1), 'violations' (list), and 'evidence' (string)

    Execution harness: Python

Evaluation tools

  • Google ADK · Orchestration Framework
  • Llama 3.3 70B · Text Analysis Model
  • Vertex AI · Platform

Recipe freeze marker

No recipe freeze time is recorded.

Frequently Asked Questions about Multimodal Flood Resilience Agent with Google ADK and Llama 3.3 70B