Personal Notes

- Before RL: 30-50 CVs (Quick Mode recommended)
  - After RL: 10-20 CVs per batch, 3-5 iterations

Estimated tokens per CV:
  - Input: ~7,000 tokens (prompts + PDF content × 2)
  - Output: ~3,000 tokens (JSON + extracted text)

  Cost per CV: ~$0.0125 (about 1.25 cents)

  ---
  Total Cost Breakdown

  Before RL (Baseline)

  | CVs | Input Cost | Output Cost | Total |
  |-----|------------|-------------|-------|
  | 30  | $0.11      | $0.27       | $0.38 |
  | 40  | $0.14      | $0.36       | $0.50 |
  | 50  | $0.18      | $0.45       | $0.63 |

  After RL (Iterative Training)

  | Batches | CVs | Input Cost | Output Cost | Total |
  |---------|-----|------------|-------------|-------|
  | 3 × 10  | 30  | $0.11      | $0.27       | $0.38 |
  | 4 × 15  | 60  | $0.21      | $0.54       | $0.75 |
  | 5 × 20  | 100 | $0.35      | $0.90       | $1.25 |

  ---
  Complete Training Pipeline Cost

  | Scenario    | Before RL      | After RL        | Grand Total |
  |-------------|----------------|-----------------|-------------|
  | Minimal     | 30 CVs ($0.38) | 30 CVs ($0.38)  | $0.76       |
  | Recommended | 40 CVs ($0.50) | 60 CVs ($0.75)  | $1.25       |
  | Thorough    | 50 CVs ($0.63) | 100 CVs ($1.25) | $1.88       |

  💡 Cost Optimization Tips

  1. Use Quick Mode: Tournament Mode (3 agents) costs 3x more (~$3.75 total)
  2. Batch processing: Your bulk_evaluation.py with --workers 3 processes efficiently
  3. Context caching: Gemini 3 Flash includes free context caching (90% savings on repeated content)
  4. Free tier: Google AI Studio includes free quota that might cover your entire training!

  Bottom line: Your entire RL training pipeline will cost $1-2 in total using Quick Mode. Very affordable! 🎯