I reach for fine-tuning only when prompts, RAG, and routing cannot hit tone or format targets — not as a first move.
In my daily workflow
- I exhaust prompt + retrieval + few-shot improvements first.
- I curate a training set from production failures and gold examples.
- I benchmark fine-tuned vs base on held-out evals, not vibes.
- I plan rollback to base model if metrics regress.
How it makes me work smarter
Fine-tuning fixes systematic format or domain style gaps. For a portfolio assistant, published writings already are the style guide — RAG usually wins. I consider fine-tuning for classifiers with stable labels, like job-fit categories, when zero-shot drifts.
My setup
- Labeled datasets from admin feedback exports
- Eval harness comparing base vs fine-tuned
- Hosted fine-tune jobs on provider with version tags
- Router flag to shadow-test new weights
On this portfolio
Job-fit classification is a candidate for fine-tuning if prompt-only approaches plateau; the analytics page collects labeled outcomes. General Q&A stays RAG-first on knowledge content.
