Skip to main content
Loading navigation
Back to challenges

Multi-Tenant Power and Token Cost Observability Pipeline with CrewAI and Gemini 2.5 Flash

Attributing real hardware power consumption (Watts), GPU VRAM usage, and dollar costs to specific tenant workflows is critical for platform providers. You will create a multi-agent telemetry pipeline using CrewAI and Gemini 2.5 Flash to parse hardware metrics, aggregate tenant token usage, and generate audited cost reports. The framework will be evaluated on strict precision of dollar-and-watt attribution across overlapping execution windows.

EngineeringHosted by Vera
Sign in free to start

What you are building

Build a CrewAI agent team powered by Gemini 2.5 Flash to ingest raw hardware telemetry metrics and attribute power and token costs to tenant workloads.

How work is evaluated

Evaluation

Evaluates CrewAI team's ability to accurately divide raw power and VRAM metrics across concurrent tenant requests.

Datasets

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

How submissions are scored

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

billing_sum_integrity

Ensures total tenant billed cost equals aggregate host power expense

Binary check

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

Dimension 2

attribution_accuracy_pct

Accuracy percentage of power attribution against analytical ground truth • target: 98 • range: 0-100

Binary check

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

Learning goals

  • Define specialized CrewAI agents (TelemetryCollector, CostCalculator, AuditReporter)

  • Integrate Gemini 2.5 Flash LLM for high-speed analysis of Prometheus hardware streams

  • Calculate proportional GPU power (Watts) and energy (kWh) per tenant based on execution duration and memory footprint

  • Produce compliant audit logs matching tenant billing SLAs

Resources

Dataset notes

Time-series GPU power logs (NVML) and concurrent tenant API request traces.

Challenge setup

Evaluates CrewAI team's ability to accurately divide raw power and VRAM metrics across concurrent tenant requests.

Challenge input

JSON with node_power_watts_series, energy_cost_per_kwh, and tenant_sessions array

CrewAI

Provides task orchestration and role-based agent collaboration.

Gemini 2.5 Flash

Offers high throughput and fast responses for dense telemetry streams.

Evaluated output

JSON detailing per_tenant_cost_usd, total_kwh_consumed, and attribution_accuracy_pct

Evaluation checks
  • Ensures total tenant billed cost equals aggregate host power expense
  • Accuracy percentage of power attribution against analytical ground truth • target: 98 • range: 0-100
Configured evaluation evidence
  • Benchmark: Green-AI-Telemetry-Bench
  • Attribution Accuracy Pct target: 98
  • 1 public reference case
Configured runtime
  • Python execution harness
  • Python sandbox (unavailable on Versalist)

Shows the first task, up to three tools, and up to three evidence items.

View technical recipe

Configured tools

Action Space
  • CrewAI · Required
  • crewAI · Optional
Policy Serving
  • Gemini 2.5 Flash · Optional

Evaluation contract

  • billing_sum_integrity · Weight 1
  • attribution_accuracy_pct · Weight 1

Metrics

  • attribution_accuracy_pct · Target 98

Benchmarks

  • Green-AI-Telemetry-Bench
  • energy

Reference cases

  • 1 public reference case

Workflow tasks

  1. Task 1

    Input: JSON with node_power_watts_series, energy_cost_per_kwh, and tenant_sessions array

    Output: JSON detailing per_tenant_cost_usd, total_kwh_consumed, and attribution_accuracy_pct

    Execution harness: Python

Evaluation tools

  • CrewAI · Agent Framework
  • Gemini 2.5 Flash · LLM Engine

Recipe freeze marker

No recipe freeze time is recorded.

Frequently Asked Questions about Multi-Tenant Power and Token Cost Observability Pipeline with CrewAI and Gemini 2.5 Flash