[P3-Research] Explore fine-tuning GPT-4o-mini for resume tailoring

[Ztrimus]

Overview

Once we have 500+ validated resume tailoring examples (input JD + original resume → output tailored resume), fine-tune GPT-4o-mini for:

Benefits

• Eliminate long system prompts → lower token costs
• More consistent output format
• Faster inference
• Domain-specific knowledge baked in

Cost estimate

• Training: ~500 examples x 1000 tokens x 3 epochs = ~$3.60
• Inference: ~2x base GPT-4o-mini price, but shorter prompts offset this

Prerequisites

• Evaluation framework ([P2-Testing] Add DeepEval evaluation metrics for resume quality scoring #57) to measure quality
• Sufficient validated examples from production usage
• A/B testing infrastructure to compare fine-tuned vs base model

Depends on

• DeepEval evaluation ([P2-Testing] Add DeepEval evaluation metrics for resume quality scoring #57)
• Instructor integration ([P1-Architecture] Integrate Instructor library for guaranteed structured LLM output #47)