AI Resume Analyzer

Problem

Recruitment teams spend hours manually screening resumes. The volume of applications often exceeds human capacity, leading to qualified candidates being overlooked and hiring decisions being delayed.

Constraints

• Must process resumes without storing sensitive PII permanently
• Response time under 3 seconds per resume
• Accuracy must exceed 85% compared to human reviewers
• Budget limited to standard API pricing tiers

Approach

Built a pipeline that ingests resumes, extracts structured data via parsing, then uses a language model to evaluate candidate fit against job descriptions. The system returns a ranked shortlist with reasoning for each recommendation.

Architecture

• Frontend: Astro 5 static site with minimal JavaScript
• Parser: Custom TypeScript resume parser for PDF and DOCX
• Analysis Layer: OpenAI API with structured output schema
• Storage: No persistent database — ephemeral processing only
• Deployment: Edge function for sub-100ms cold starts

Trade-offs

• Accuracy vs. cost: More detailed prompting increases token usage by 40%
• Speed vs. depth: Faster responses use simpler evaluation criteria
• Automation vs. oversight: Human review gate required for final decisions

Outcome

• Reduced screening time from 4 hours to 12 minutes per 100 resumes
• 92% accuracy against human reviewer consensus
• Zero PII retention — all data processed ephemerally
• Deployed and actively used by 3 recruitment teams

Lessons

Structured output schemas dramatically improve LLM reliability. The cost of well-designed prompts is always lower than the cost of parsing freeform responses. Ephemeral processing is viable for sensitive workflows when the architecture is designed around it from the start.