Build the application

## Build
Correlate per-request batch tokens with measured accelerator power sensor logs. Synthesize carbon footprint summaries and efficiency improvement plans using hosted Claude.

## User
Datacenter Sustainability Lead

## Workflow
1. Ingest normalized time-series power telemetry from GPU power sensors and batch serving token counts.
2. Compute joules-per-token and convert total watt-hours to carbon equivalents using local regional grid emission factors.
3. Invoke hosted Claude to identify power-inefficient batching patterns and draft optimization proposals.
4. Produce an energy ledger report detailing departmental carbon footprints and recommended power caps.

## Interface
Static HTML sustainability report generator and downloadable CSV carbon allocation ledger.

## Data
Normalized power telemetry logs (Watts) paired with synthetic continuous batching request logs containing input/output token counts.
Starter examples below are synthetic. They are not downloaded records.

## Source
[NVIDIA DCGM exporter](https://raw.githubusercontent.com/NVIDIA/dcgm-exporter/main/README.md)
Utilizes DCGM exporter power consumption metrics (DCGM_FI_DEV_POWER_USAGE) to calculate physical energy draw.

## Tools
evidently-ai, anthropic.
Evidently AI monitors data drift and distribution changes in power-per-token consumption over time.
Hosted a currently available hosted Claude model accessed via Anthropic Python SDK; purely analytical role.

## Deliverables
Submit runnable source, pinned dependencies, setup commands, a working interface, example outputs, and a recorded end-to-end demo. Include tests and an action log.

## Acceptance
- Calculates exact watt-hours and joules-per-token metrics from telemetry.
- Applies regional emission factors accurately to produce kgCO2e values.
- Synthesizes actionable efficiency advice via hosted Claude without fabricated metrics.

## Starter Examples
```json
[
  {
    "input": {
      "team": "SearchPlatform",
      "tokens_processed": 5000000,
      "duration_hours": 10,
      "avg_power_watts": 350,
      "grid_carbon_intensity_g_kwh": 380
    },
    "expected_output": {
      "total_kwh": 3.5,
      "joules_per_token": 2.52,
      "total_emissions_kg_co2": 1.33,
      "sustainability_advice": "Joules per token are within standard bounds. Recommend dynamic power capping during low-traffic windows."
    }
  },
  {
    "input": {
      "team": "DocumentSummarizer",
      "tokens_processed": 50000,
      "duration_hours": 8,
      "avg_power_watts": 320,
      "grid_carbon_intensity_g_kwh": 400
    },
    "expected_output": {
      "total_kwh": 2.56,
      "joules_per_token": 184.32,
      "total_emissions_kg_co2": 1.02,
      "sustainability_advice": "Extremely high joules per token detected due to low batch occupancy. Recommend request queue pooling to increase continuous batch utilization."
    }
  }
]
```

## Safety
Calculations rely solely on explicitly provided telemetry; does not enforce hardware power caps directly.
Use mock external writes. Require human review before any real deployment.