CV Tailor
Agentic AI - Document Generation - Authenticated SaaS
An agentic application that tailors a CV to each job. It maps every requirement in a job description to evidence in your CV, rewrites only what the evidence supports, and tracks each application through to the outcome.
Problem
AI CV tools tend to please the job description rather than tell the truth: they add skills you do not have and invent numbers. A CV with fabricated claims is a liability the moment an interviewer asks about it.
Approach
Evidence first, rewriting second. Each job requirement is matched against the CV and rated strong, transferable or not found. Only evidenced requirements shape the rewrite; the rest are shown to the user as gaps. Validators then enforce the rules in code: numbers must exist in the original, and employers, titles and dates pass through unchanged. The user approves each change before anything is exported.
Trade-off
I enforce honesty with validators after generation instead of trusting the prompt. That means more checks and occasional retries, and a full tailoring run takes minutes rather than seconds. In exchange, an invented claim cannot reach an exported document.
Architecture
Full-stack pipeline: FastAPI + Supabase (auth, storage, Postgres) behind a Next.js/TypeScript app. CVs are ingested from DOCX, PDF or pasted text into an editable document. Job descriptions come in as pasted text or a Greenhouse/LinkedIn link. A free keyword check runs in seconds; full tailoring researches the employer, scores fit, and rates each requirement as strong, transferable or not found. Rewrites pass validation before anything is shown: figures are checked against the original CV, and employers, titles and dates are copied through untouched. Every change needs approval before export to DOCX and PDF, alongside a cover letter. Each role is stored separately with a field-by-field comparison against the original and an outcome tracker. Operations: structured LLM output with strict JSON schema and a prompt-engineered fallback, multi-provider routing with retries and fallback, per-call telemetry (prompt, response, latency, tokens) in Postgres, request-ID logging, rate limiting and a monthly tailoring quota.
Results
Each application gets its own tailored CV and cover letter, stored separately and compared field by field against the original, with status tracked from applied through interview to offer. Per-call LLM telemetry in Postgres makes every prompt, response, latency and provider fallback inspectable after the fact.
Highlights
- Requirement-to-evidence mapping; gaps are reported, never written in
- Invented metrics rejected automatically against the original CV
- Every change approved by the user before DOCX/PDF export
Technologies
- Next.js
- TypeScript
- FastAPI
- Supabase
- PostgreSQL
- Tailwind CSS
- Vercel
Screenshots
