BREV A100 batch classification — 10.5M digital twins

{
  "cost": {
    "input_tokens": 118,
    "estimated_usd": 0.002933,
    "output_tokens": 958
  },
  "model": "gemini-3-flash-preview",
  "status": "success",
  "transport": "gemini_api",
  "latency_ms": 10599,
  "response_text": "This plan outlines the execution strategy for classifying the 10.5M records using the BREV A100 cluster and the 5-tier cascade logic.\n\n### 1. Environment Verification\nEnsure your environment points to the ATLAS vLLM endpoint and the `data_job` MCP is active.\n\n```bash\n# Verify vLLM Endpoint\ncurl http://localhost:8000/v1/models\n\n# Initialize data_job tracking\nmcp data_job start --job-name \"BREV_HTS_CLASSIFICATION_10.5M\"\n```\n\n---\n\n### 2. Execution Phase 1: The Dry Run\n**Purpose:** Validate database connectivity, schema mapping, and cascade logic without consuming credits or writing to the DB.\n\n```bash\npython batch_classifier.py \\\n  --dry-run \\\n  --source \"product_digital_twins\" \\\n  --where \"hts_primary IS NULL\" \\\n  --cascade \"noise,oracle,vllm,haiku,opus\" \\\n  --vllm-url \"http://localhost:8000/v1\"\n```\n*   **Check for:** SQL syntax errors, \"Noise\" filter regex accuracy, and Oracle KB connection strings.\n\n---\n\n### 3. Execution Phase 2: Pilot (Limit 100)\n**Purpose:** Measure the \"fall-through\" rate. If vLLM handles >90%, costs stay low. If too many hit Opus, the $36 budget will be exceeded.\n\n```bash\npython batch_classifier.py \\\n  --limit 100 \\\n  --source \"product_digital_twins\" \\\n  --where \"hts_primary IS NULL\" \\\n  --output \"mcp_tracking\" \\\n  --commit\n```\n*   **Monitor:** `data_job` dashboard for the distribution of classification sources (e.g., `source: vllm` vs `source: opus`).\n\n---\n\n### 4. Execution Phase 3: Full Production Run\n**Target:** ~10.5M records.\n**Estimated Time:** 20 Hours (approx. 145 records per second across the parallel cascade).\n\n```bash\nnohup python batch_classifier.py \\\n  --source \"product_digital_twins\" \\\n  --where \"hts_primary IS NULL\" \\\n  --batch-size 1000 \\\n  --concurrency 50 \\\n  --mcp-track \\\n  --commit > production_run.log 2>&1 &\n```\n\n---\n\n### 5. The 5-Tier Cascade Logic\nThe script will process each record through this hierarchy to optimize cost and accuracy:\n\n1.  **Noise Filter:** Immediate rejection of \"test\", \"abc\", or null descriptions.\n2.  **Oracle KB:** Exact/Fuzzy match against existing `hts_primary` mappings in the master knowledge base.\n3.  **vLLM (ATLAS):** Local A100 inference. Primary engine for 95% of standard products.\n4.  **Claude 3 Haiku:** Triggered if vLLM confidence < 0.7 or ambiguous classification.\n5.  **Claude 3 Opus:** Triggered for complex chemical compounds or multi-component machinery requiring deep reasoning.\n\n---\n\n### 6. Monitoring & Telemetry\nUse the **data_job MCP** to track progress in real-time.\n\n*   **Progress Check:** `mcp data_job status --job-id [JOB_ID]`\n*   **Cost Tracking:** \n    *   *Local (vLLM):* $0 (Compute already paid)\n    *   *API (Haiku/Opus):* Monitor `usage_limit` in the script logs.\n*   **Success Metrics:**\n    *   `throughput_rpm`: Records per minute.\n    *   `hit_rate_vllm`: Percentage handled by ATLAS.\n    *   `null_reduction`: Count of `hts_primary` fields populated.\n\n### Critical Alerts\n*   **OOM on A100:** If vLLM crashes, the script is configured to pause and retry the connection before falling back entirely to Haiku.\n*   **Rate Limiting:** If Opus hits tier limits, the script will throttle the high-tier cascade while maintaining vLLM throughput."
}