Build an Accelerator Chassis Power Budget and Throttling Reviewer — implementation

## Build
Correlate server chassis electrical power caps with per-device DCGM utilization logs. Formulate dynamic power balancing schedules and throttling prevention plans using hosted Claude.

## User
Datacenter Power and Facilities Engineer

## Workflow
1. Ingest rack PDU power limits and per-accelerator DCGM power consumption telemetry using pandas.
2. Detect electrical headroom deficits where total peak draw risks tripping rack-level circuit breakers.
3. Invoke hosted Claude via Anthropic API to synthesize power cap balancing proposals across non-critical workers.
4. Generate a structured power allocation manifest detailing per-GPU watt limits validated via pydantic.

## Interface
Dashboard visualizing rack power envelope utilization and Claude-recommended power capping profiles.

## Data
Simplified synthetic app records representing power telemetry logs from server PDUs and DCGM metrics across 8-accelerator chassis.
Starter examples below are synthetic. They are not downloaded records.

## Source
[NVIDIA DCGM exporter](https://raw.githubusercontent.com/NVIDIA/dcgm-exporter/main/README.md)
Extracts DCGM power metrics (DCGM_FI_DEV_POWER_USAGE) and hardware power limit registers.

## Tools
dspy, anthropic, pandas, pydantic.
DSPy compiles robust prompt pipelines to guarantee hosted Claude produces schema-compliant power allocation limits.
Hosted a currently available hosted Claude model accessed via official SDK; does not emit live electrical breaker or firmware commands.

## 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
- Identifies potential PDU overload conditions when peak sum exceeds rack threshold.
- Calculates balanced power allocation across active server chassis.
- Produces hosted Claude explanations for selected power capping setpoints.

## Starter Examples
```json
[
  {
    "input": {
      "pdu_max_kw": 12,
      "chassis_count": 4,
      "unconstrained_peak_kw": 14.5,
      "priority_chassis_index": 0
    },
    "expected_output": {
      "breaker_trip_risk": true,
      "recommended_chassis_caps_kw": [
        4,
        2.6,
        2.6,
        2.6
      ],
      "total_budget_kw": 11.8,
      "allocation_rationale": "Chassis 0 maintained at peak demand for SLA compliance; remaining chassis capped to prevent PDU overload."
    }
  },
  {
    "input": {
      "pdu_max_kw": 12,
      "chassis_count": 4,
      "unconstrained_peak_kw": 9.2,
      "priority_chassis_index": 0
    },
    "expected_output": {
      "breaker_trip_risk": false,
      "recommended_chassis_caps_kw": [
        3.5,
        2.5,
        2,
        1.2
      ],
      "total_budget_kw": 9.2,
      "allocation_rationale": "Operating load is well within the 12 kW PDU envelope. No power capping required."
    }
  }
]
```

## Safety
Provides advisory power capping allocations; does not issue live IPMI power cap updates.
Use mock external writes. Require human review before any real deployment.