By NotGhosty Engineering Team — Published on Tech, Open Source, and Candidate Privacy
When job seekers upload their CV to modern "free" ATS resume checkers, they assume they are receiving a simple diagnostic utility.
Behind the scenes, the reality is far more predatory: most commercial resume checkers persist your raw PDF/Word documents in cloud object storage (S3/GCS) and extract your phone number, email, work history, and home address into candidate databases. These databases are routinely packaged and sold to third-party headhunters, data brokers, or used to train proprietary recruitment LLMs without explicit consent.
When building NotGhosty, we engineered our parser under a non-negotiable architectural invariant: Strictly Zero Storage. 100% Volatile Memory Execution.
Here is a technical breakdown of how we built an in-memory ATS parser using Python 3.12, pypdf, and python-docx that diagnoses layout traps and immediately purges raw file bytes from RAM.
The Zero-Storage In-Memory Pipeline
Rather than saving incoming files to a local /tmp directory or an S3 bucket, our FastAPI endpoint ingests documents as raw bytes directly into an in-memory io.BytesIO buffer:
import io import pypdf import docx from typing import Dict, Any, Tuple class ATSParserService: @classmethod def parse_document(cls, file_bytes: bytes, filename: str) -> Dict[str, Any]: filename_lower = filename.lower() if filename_lower.endswith(".pdf"): text, warnings = cls._parse_pdf(file_bytes) elif filename_lower.endswith((".docx", ".doc")): text, warnings = cls._parse_docx(file_bytes) else: text = file_bytes.decode("utf-8", errors="ignore") warnings = [] # Structured Entity Extraction full_name = cls._extract_name(text) email = cls._extract_email(text) phone = cls._extract_phone(text) skills = cls._extract_skills(text) experience = cls._extract_experience(text) education = cls._extract_education(text) # Raw file_bytes drop out of scope here and are freed by Python GC return { "name": full_name, "email": email, "phone": phone, "skills": skills, "experience": experience, "education": education, "warnings": warnings } 1. In-Memory PDF Extraction with Trap Detection
Enterprise ATS systems (Workday, Taleo, Greenhouse) choke on multi-column layouts, embedded tables, and vector graphics. Our in-memory stream analyzes layout traps on the fly:
@classmethod def _parse_pdf(cls, file_bytes: bytes) -> Tuple[str, List[Dict[str, str]]]: warnings = [] text_chunks = [] stream = io.BytesIO(file_bytes) try: reader = pypdf.PdfReader(stream) if reader.is_encrypted: warnings.append({ "category": "Format", "severity": "critical", "message": "PDF is password protected or encrypted.", "recommendation": "Export an unencrypted standard PDF." }) return "", warnings for page_idx, page in enumerate(reader.pages): page_text = page.extract_text() or "" text_chunks.append(page_text) # Detect multi-column layout trap via tab and whitespace heuristics lines = page_text.splitlines() column_gaps = sum(1 for line in lines if " " in line or "\t" in line) if column_gaps > len(lines) * 0.35: warnings.append({ "category": "Layout Trap", "severity": "warning", "message": f"Page {page_idx + 1} exhibits multi-column spacing.", "recommendation": "Use a single-column linear layout to prevent ATS text scrambling." }) finally: stream.close() del file_bytes # Explicit garbage collection trigger return "\n".join(text_chunks), warnings 2. Diagnosing Common ATS Parsing Failures
When resumes fail to pass corporate ATS screens, it is rarely due to a lack of qualifications. It is almost always structural:
| Trap Type | Real-World Impact | NotGhosty Diagnostic Rule |
|---|---|---|
| Multi-Column Layouts | Scanners read horizontally across columns, interleaving job titles with unrelated bullet points. | Flags lines with high whitespace variance and tab jumps. |
| Fancy Graphic Headers | Contact details placed inside SVG/image headers are completely invisible to text extractors. | Validates presence of valid email regex and E.164 phone tokens in the top 20% of text. |
| Stylized Section Titles | Headers like "Where I've Been" or "Core Superpowers" are categorized as unknown text. | Compares section headers against standard dictionaries (Experience, Education, Skills). |
| Embedded Canva Tables | Cell borders flatten into single line streams, scrambling dates and company names. | Scans DOCX XML tags (w:tbl) and flags table-dependent layouts. |
3. Dynamic On-Demand Clean CV Generation
When a candidate's resume triggers critical layout warnings, NotGhosty doesn't just display errors—it compiles an optimized, single-column ATS-certified PDF on the fly using reportlab.
The generated PDF strictly adheres to ATS best practices:
- Standard 0.5-inch margins
- Standard Helvetica / Times typeface hierarchy
- Linear top-to-bottom reading stream
- UTF-8 clean character encoding without stylized glyphs
Key Takeaway: Privacy by Design
Privacy is not a policy written in terms of service; it is an architectural decision.
By handling all byte parsing in volatile memory buffers and enforcing immediate dereferencing, candidates can diagnose and optimize their resumes without creating digital breadcrumbs.
- Test your resume's parsing readability anonymously at https://notghosty.com/ats.
- Check the credibility of job postings in real time at https://notghosty.com.
Originally researched and published by the team at NotGhosty.

