Build the application

## Build
Parse DCGM link telemetry during burn-in. Validate empirical cluster topology against hardware manifests and generate acceptance reports using hosted Claude to identify miswired or underperforming interconnects.

## User
Cluster Infrastructure Hardware Engineer

## Workflow
1. Ingest initial NVLink CRC rates and PCIe bandwidth telemetry from newly racked cluster nodes.
2. Compare empirical communication latency and link matrices against the expected hardware procurement manifest.
3. Query hosted Claude API to synthesize a burn-in acceptance report identifying miswired links.
4. Output a cluster acceptance JSON payload to unblock inventory provisioning or trigger recabling.

## Interface
Terminal CLI and automated JSON report generator for cluster provisioning.

## Data
Synthetic simplified app records of DCGM fabric error logs, PCIe AER records, and JSON hardware procurement manifests.
Starter examples below are synthetic. They are not downloaded records.

## Source
[NVIDIA DCGM exporter](https://raw.githubusercontent.com/NVIDIA/dcgm-exporter/main/README.md)
Uses DCGM exporter telemetry counters to validate NVLink error metrics and PCIe throughput against manifests.

## Tools
pydantic-ai, anthropic-claude, fastapi.
Validates structured multi-GPU topology link matrices and enforces acceptance schema definitions.
Hosted Claude API for burn-in diagnostics; strictly remote API calls.

## 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 specific GPU-to-GPU link pairs failing bandwidth expectations during burn-in.
- Distinguishes between manifest mismatches and physical connector degradation.
- Outputs concrete hardware remediation steps (e.g., recable NVLink topology).
- Produces valid acceptance validation JSON payloads without live restarts.

## Starter Examples
```json
[
  {
    "input": {
      "node_name": "compute-dgx-new-14",
      "expected_manifest_bandwidth_gbps": 450,
      "gpu_pairs": [
        {
          "gpu_a": 0,
          "gpu_b": 1,
          "nvlink_crc_errors": 0,
          "bandwidth_gbps": 450
        },
        {
          "gpu_a": 2,
          "gpu_b": 3,
          "nvlink_crc_errors": 184,
          "bandwidth_gbps": 75.2
        }
      ]
    },
    "expected_output": {
      "node_status": "FAILED_BURN_IN",
      "impacted_link": "GPU_2_TO_GPU_3",
      "error_type": "BANDWIDTH_MISMATCH",
      "technician_repair_plan": [
        "Inspect physical NVLink mezzanine connector between socket 2 and socket 3.",
        "Clean optical pins and reseat link module prior to re-running burn-in suite."
      ],
      "severity": "BLOCKING"
    }
  },
  {
    "input": {
      "node_name": "compute-dgx-new-15",
      "expected_manifest_bandwidth_gbps": 450,
      "gpu_pairs": [
        {
          "gpu_a": 0,
          "gpu_b": 1,
          "nvlink_crc_errors": 0,
          "bandwidth_gbps": 450
        },
        {
          "gpu_a": 1,
          "gpu_b": 2,
          "nvlink_crc_errors": 0,
          "bandwidth_gbps": 450
        }
      ]
    },
    "expected_output": {
      "node_status": "ACCEPTED",
      "impacted_link": null,
      "error_type": "NONE",
      "technician_repair_plan": [],
      "severity": "NONE"
    }
  }
]
```

## Safety
Read-only telemetry and manifest analysis. Does not perform physical power-cycling.
Use mock external writes. Require human review before any real deployment.