cv-tailoring

Pass master skill and project blocks plus the job description to the LLM with a prompt contract forbidding new content. Validate the returned JSON with a Pydantic schema containing tailored_summary, selected_project_ids, top_skill_ids, and keyword_injections.

Extract keywords from each CV version's metadata and ask the LLM to compare them against the job's detected stack. The model returns only the ID of the best-matching CV version as JSON.

Use an immutable block selection approach where the LLM only selects and reorders existing blocks by ID. The system prompt explicitly forbids generating new data or modifying existing sentences, and only the summary may be rewritten.

fpdf2 generates cover letter PDFs via the FPDF class by adding pages and writing tailored content. Reportlab is an alternative mentioned for the same task, with output paths stored in the applications table.

Use claude-sonnet-4 as the primary model since language quality matters for CVs. Configure a fallback chain through gpt-4o-mini, groq/llama-3-70b, and ollama/llama3:8b for rate-limit resilience.

Avoid it for general prompt engineering unrelated to CV content or for managing CV files on disk and S3, which is infrastructure work. It also cannot create genuinely new experience entries, only reorganize existing ones.