Build the application

## Build
Analyze DCGM memory error telemetry using pandas to detect uncorrectable ECC faults and row remapping saturation on local GPUs. Synthesize maintenance dossiers using Pydantic AI and Claude to order physical hardware replacement.

## User
Datacenter Hardware Reliability Technician

## Workflow
1. Ingest Prometheus-formatted DCGM time-series telemetry streams containing ECC error counts and row remapping statuses into pandas DataFrames.
2. Isolate GPUs exhibiting accelerating multi-bit ECC uncorrectable errors or nearing row remapping limits.
3. Query hosted Claude API via Pydantic AI to synthesize a physical triage guide for hardware replacement.
4. Generate a structured maintenance packet JSON with technician action items and simulated RMA work-order stubs.

## Interface
CLI tool that outputs formatted terminal inspection briefs and writes signed maintenance JSON packets.

## Data
Simulated DCGM CSV/Prometheus metric logs in a local fixtures folder mocking ECC error counts and row remap statuses.
Starter examples below are synthetic. They are not downloaded records.

## Source
[NVIDIA DCGM exporter](https://raw.githubusercontent.com/NVIDIA/dcgm-exporter/main/README.md)
References DCGM exporter Prometheus telemetry metrics for ECC memory errors, volatile single-bit, and uncorrectable double-bit errors.

## Tools
pydantic-ai, anthropic, pandas.
Ensures strict type-safe validation of DCGM telemetry inputs and technician maintenance packet schemas.
Hosted Claude API for technical synthesis and root-cause explanation; no local weights.

## 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
- Correctly flags GPUs requiring replacement when multi-bit ECC errors exceed thresholds.
- Outputs technician checklist referencing exact physical chassis slots.
- Rejects input payloads with corrupted timestamps or missing GPU identifiers.
- Emits complete JSON maintenance packet without executing actual hardware BMC commands.

## Starter Examples
```json
[
  {
    "input": {
      "node_id": "rack-4-host-02",
      "gpu_index": 3,
      "telemetry": [
        {
          "timestamp": 1710000000,
          "ecc_db_error": 12,
          "row_remap_pending": true
        }
      ]
    },
    "expected_output": {
      "packet_id": "MAINT-R4-H02-G3",
      "severity": "CRITICAL",
      "issue_summary": "Uncorrectable ECC double-bit errors detected.",
      "technician_actions": [
        "Initiate node drain.",
        "Replace GPU module 3."
      ],
      "simulated_ticket_status": "QUEUED_FOR_RMA",
      "note": "Simplified synthetic app record"
    }
  },
  {
    "input": {
      "node_id": "rack-4-host-02",
      "gpu_index": 1,
      "telemetry": [
        {
          "timestamp": 1710000000,
          "ecc_db_error": 0,
          "row_remap_pending": false
        }
      ]
    },
    "expected_output": {
      "packet_id": "MAINT-R4-H02-G1",
      "severity": "NORMAL",
      "issue_summary": "Memory operating nominally.",
      "technician_actions": [],
      "simulated_ticket_status": "NO_ACTION_REQUIRED",
      "note": "Simplified synthetic app record"
    }
  }
]
```

## Safety
Telemetry is read-only. Maintenance packets simulate ticketing without modifying BMC or server power states.
Use mock external writes. Require human review before any real deployment.