Over 97% of Fortune 500 companies use Applicant Tracking Systems to filter resumes before a human ever sees them. These systems reject up to 75% of applicants automatically based on keyword matches, section parsing, and formatting compatibility. Your qualifications don't matter if the software can't read them.
ATS Boss uses GPT-5-mini's reasoning capabilities (OpenAI o4-mini) to simulate how Workday, Greenhouse, and Ashby would actually parse and score your resume. This isn't keyword counting — the model reasons step-by-step through each system's documented parsing logic, scoring weights, and ranking criteria to produce an analysis that mirrors real ATS behavior.
Upload your resume, paste the job description, and get a detailed scoring breakdown showing exactly where you stand — which keywords matched, which were missed, which sections the parser found or skipped, and where you'd land in the recruiter's queue. Then generate an optimized PDF resume that addresses every issue found, using only your real experience.
How It Works
Deep ATS Simulation
GPT-5-mini (o4-mini) is prompted with the exact parsing rules, scoring weights, and ranking logic documented for each ATS system. It reasons through your resume step-by-step — detecting sections, extracting keywords, checking formatting — and produces a scored analysis with specific issues and fixes.
Optimized Resume PDF
GPT-4o-mini takes your original resume, the job description, and the full Phase 1 analysis — scoring breakdown, near-misses, missing keywords, critical issues — and restructures your content to fix every identified problem. The output is rendered into a single-page PDF using ReportLab with ATS-specific formatting.
ON HONESTY: The generated resume only uses information from your original resume. It will never fabricate metrics, invent skills you don't have, or add experience that isn't yours. Missing keywords are incorporated only by rephrasing your existing experience to use the job description's terminology — not by making things up.
Supported ATS Systems
Workday
Exact string matching. No synonym understanding. Standard headings only. The strictest of the three — if your resume doesn't use the JD's exact phrases, Workday won't find them.
SCORING WEIGHTS:
WHAT WE SIMULATE:
- →Exact keyword matching (case-insensitive, no synonyms)
- →Standard section heading detection and skipping
- →Single-column layout verification
- →Near-miss identification (why a keyword almost matched)
Greenhouse
Semantic matching with structured data extraction. More forgiving on synonyms but strict on data quality — dates, titles, and fields must be cleanly parseable.
SCORING WEIGHTS:
WHAT WE SIMULATE:
- →Semantic keyword matching with confidence scores
- →Structured field extraction (company, title, dates)
- →Data extraction quality assessment
- →Experience-to-JD alignment scoring
Ashby
AI-powered matching focused on impact. Cares less about keywords, more about quantified achievements, demonstrated skills, and career trajectory.
SCORING WEIGHTS:
WHAT WE SIMULATE:
- →Quantified achievement extraction and scoring
- →Skill inference from experience context
- →Career progression and trajectory analysis
- →Standout factor identification
Under the Hood
This is a real engineering project, not a wrapper around a single API call. Here's what's actually running when you hit "Analyze."
GPT-5-mini (o4-mini)
Each ATS system has a dedicated analyzer with system-specific prompts that encode the documented parsing rules, scoring weights, and ranking thresholds. The model uses extended thinking to reason through section detection, keyword extraction, formatting analysis, and scoring calculation step by step.
GPT-4o-mini + ReportLab
The full analysis — scoring breakdown, near-misses, missing keywords, critical issues, ATS-specific data — is fed to GPT-4o-mini to restructure your resume content. The structured JSON output is then rendered into a single-page PDF with dynamic spacing that fills the entire page.
Anti-Hallucination Guardrails
The generator is constrained to only use information from your original resume. Multi-layered prompt rules prevent the model from fabricating metrics, inventing skills, or adding experience that doesn't exist. Keywords are integrated by rephrasing existing bullets — not by making new claims.
FastAPI + Next.js
Python backend with FastAPI serving the analysis and generation endpoints. Next.js 14 frontend with App Router. PDF rendering with ReportLab. Full cost transparency shown for every API call — you can see exactly what each analysis costs to run.
Want to see how this works internally?
This is an open engineering project. If you're curious about the prompt engineering, the scoring simulation, the PDF generation pipeline, or want to collaborate — let's connect.