I Built a Zero-Trust Resume Pipeline to Stop AI from Hallucinating

A developer built EigenCV, a zero-trust resume pipeline that uses Infrastructure-as-Code to generate ATS-optimized LaTeX resumes without allowing AI to hallucinate skills. The system treats career history as an immutable database and crashes the build if the AI attempts to add unverified information. EigenCV is available as an open-source repository that can run locally or inside ChatGPT.

Stop letting ChatGPT hallucinate skills you don't have. A production-grade Infrastructure-as-Code IaC pipeline for generating ATS-perfect, highly tailored LaTeX resumes without sacrificing your integrity. The Industry Standard Commercial AI Builders : You tell an AI to "optimize my resume for this job." The AI treats your resume as a creative writing exercise. It quietly hallucinates skills, inflates job titles, and paraphrases your engineering achievements into generic HR buzzwords. The result is a PDF that beats the ATS but fails the technical interview because it's full of lies. The EigenCV Approach Zero-Trust : We treat your career history as an immutable database. The AI is strictly an orchestration layer . It does not write your resume; it queries your database to pull the most relevant, pre-verified bullet points. If the AI attempts to go rogue and hallucinate a missing skill into your profile to artificially boost your ATS score, the Python compiler's Lie Detector intercepts it and hard-crashes the build. Obvious lies never make it into the PDF. The sample above is just one configuration. EigenCV puts you in complete control: Dynamic Sections: Not all sections are mandatory. Choose exactly what to include and reorder them instantly e.g., move 'Education' to the top by simply editing the comma-separated cvorder variable. Change Layouts: Seamlessly swap between professional LaTeX templates e.g., Awesome-CV , EigenCV-Modern . Brand Yourself: Inject custom corporate accent colors to match the company you're applying to.- Multilingual: Generate applications in English, German, or any other language using the built-in translation matrix. You don't need to know Python, LaTeX, or Git to use EigenCV. You can run the entire pipeline directly in ChatGPT. Download this entire repository as a ZIP file. Build your Master Database: Upload the ZIP to ChatGPT requires Advanced Data Analysis or Claude, along with ALL your old resumes. Tell the AI: "Extract all my career facts from these resumes and populate the EigenCV JSON database." Apply to a Job: Paste the Job Description and tell the AI: "Please apply to this job using the instructions found in docs/AI CLOUD PROMPT.md ." Download your PDF: ChatGPT will automatically match your database to the job, generate the JSON, and run the chatgpt run.py wrapper script. Because we include a dedicated "Cloud-Safe" LaTeX template, ChatGPT will render the PDF directly in its sandbox and give you a download link Fallback: If ChatGPT times out, it will still generate the .tex code. You can drag & drop that code into a free Overleaf account for instant rendering or render the pdf locally using the pdflatex . If you prefer terminal workflows, need maximum build speed, or require strict data privacy, run the pipeline locally. Prerequisites: You need Python 3.11+ and a LaTeX distribution e.g., TeX Live or MiKTeX . Alternatively, simply open this repo in VS Code and click "Reopen in Container" to use our pre-built Docker environment Install Dependencies: Run pip install -r requirements.txt . Agent Setup: For Maximum Speed: Open the repository using an Agentic CLI like Antigravity / Claude Code or an Agentic IDE Cursor / Windsurf connected to their standard cloud APIs. For 100 % Hardcore Privacy: Point your Agentic IDE to a local model like Ollama or LM Studio so your data NEVER leaves your machine. Build the DB: Tell the Agent: "Migrate my old CV. Follow to build your Zero-Trust database. AI START HERE.md ." Apply: Paste a Job Description and say: "Apply to this job. Follow AI START HERE.md ." Automation: The Agent will automatically route the prompts, generate the strict JSON, and execute the Python scripts locally to render your PDF and calculate your ATS score Most AI tools use a "generate and pray" approach. EigenCV uses Agentic Determinism . Here is how we guarantee a flawless, hallucination-free application: The Immutable Master Database: You don't paste your resume into a chat window. Your career history lives offline as a structured JSON/Markdown database on your hard drive. Every achievement, project, and skill has a mathematically verifiable ID e.g., proj aws migration . AI Orchestration The Brain : When applying for a job, the AI reads the Job Description and acts as a strategic orchestrator. Instead of writing text, it executes a precise query against your database, returning only the IDs that maximize your ATS match score. It crafts a hyper-authentic Cover Letter based exclusively on your verified profile data. Deterministic Compilation The Muscle : The AI is entirely locked out of the rendering process. The EigenCV Python compiler takes the approved list of IDs, fetches the exact text from your offline database, and deterministically injects it into a stunning, ATS-optimized LaTeX template. Flawless layouts, no hallucinations. sequenceDiagram autonumber actor U as 👤 User participant LLM as 🧠 Agent participant DB as 🗄️ Database participant Py as 🐍 Python Tools rect rgb 40, 40, 60 Note over U,DB: Phase 0: Setup & Privacy U- Py: Run scrub data.py Py- DB: Wipe private data U- LLM: Upload old CV LLM- DB: Build JSON Database end rect rgb 20, 40, 20 Note over U,Py: Phase 1: Routing U- LLM: Provide Job Description LLM- DB: Query database DB-- LLM: Return valid IDs LLM- LLM: Generate build config.json end rect rgb 60, 40, 20 Note over LLM,Py: Phase 2: Compile LLM- Py: Execute cv compiler.py Note over Py,DB: 🏥 Healer auto-corrects IDs<br 🛡️ Lie Detector prevents hallucinations Py- DB: Fetch verified text Py-- LLM: Render PDF via Jinja2 end rect rgb 40, 60, 40 Note over LLM,Py: Phase 3: Verification LLM- Py: Execute check ats score.py Py- Py: Parse compiled PDF Py-- LLM: Return ATS Score LLM-- U: Present Final PDF & Score end +----------------------------- EigenCV Compiler ------------------------------+ | Compiling CV from build config.json... | +-----------------------------------------------------------------------------+ Using layout template: eigencv resume.tex.j2, Locale: en +--------------------------- Zero-Trust Violation ----------------------------+ | Zero-Trust Violation: You declared 'Rust' as a missing skill, but it was | | hallucinated into the CV output | | You cannot artificially inject skills you do not have into free-text fields.| +-----------------------------------------------------------------------------+ ValueError: Hallucinated skill detected: Rust The user removes the hallucinated skill and recompiles: +----------------------------- EigenCV Compiler ------------------------------+ | Compiling CV from build config.json... | +-----------------------------------------------------------------------------+ Successfully compiled CV to CV-Applicant-Google.tex Auto-compiling PDFs with pdflatex... Successfully compiled CV-Applicant-Google.pdf ATS Match Score: 85.0 % +-------------------------------------------------------------------+ | Category | Skills | |-----------------+-------------------------------------------------| | Missing 1 | Rust | +-------------------------------------------------------------------+ ATS Penalty Applied: 1 critical gap identified. - 🛡️ Zero-Trust & The EigenTruth Engine: Your career history lives in a static JSON database. If the LLM attempts to hallucinate a skill you don't have into your profile to artificially boost your ATS score, the compiler's EigenTruth Engine our Lie Detector catches it and hard-crashes the build. - 🔒 Immutable Database: Your bullet points and skills are strictly IMMUTABLE . You can maintain them yourself or use LLMs to prep them, but within the EigenCV pipeline, the AI is only allowed to select them, never rewrite them. - ✍️ Authentic Cover Letters: The AI uses your personal dossier.md to write hyper-authentic cover letters based only on your real soft skills and hobbies, eliminating corporate fluff. - 🎨 Corporate Auto-Coloring: The AI automatically deduces the target company's corporate identity and dynamically injects matching accent colors into the LaTeX output or you can override it manually . - 🧮 Advanced ATS Engine & Reality Checks: The post-compilation Python parser calculates a mathematically honest ATS keyword match score. Meanwhile, the AI Agent acts as a ruthless filter, estimating interview/offer probabilities and salary ranges based strictly on your verified skills, creating a realistic Probability Matrix. - 🏥 Self-Healing IDs: If the LLM makes a minor typo when selecting an ID from your database e.g., aws mig instead of aws migration , the compiler's built-in rapidfuzz heuristics will auto-heal the ID, preventing brittle build crashes. - 📄 Automated LaTeX Compilation: No more broken LaTeX parsing or missing brackets. The AI generates a strictly typed Pydantic JSON schema, deterministically compiled into beautiful Jinja2 LaTeX templates. - 🌍 Multi-Language Support & Auto-Translation Beta : Applying abroad? The system supports native multi-language CVs with strict language mismatch prevention, and features an experimental auto-translation engine to dynamically localize your database. - 🏗️ Dynamic Section Routing: Don't have any open-source projects for a specific application? Simply omit the array in the JSON. The Jinja2 engine will dynamically hide the section and recalculate the LaTeX geometry without leaving awkward whitespace. - 🐳 Containerized Reproducibility: Comes with a pre-configured VS Code DevContainer. Boot a fully sandboxed environment to get a full TeX Live distribution inside Docker. Note: The initial Docker build downloads the 4GB distribution, grab a coffee. After that, compile CVs locally without polluting your host machine. - 🕵️ 100 % Local & Privacy-First Optional : Your career data never leaves your machine unless you explicitly send it to an LLM via your trusted API or Agent. No web services, no data harvesting. EigenCV prioritizes deterministic reliability over trendy AI complexity. For the engineers reading this, here is a transparent look at the core architectural decisions behind the pipeline: We do not use Vector Databases Chroma, Pinecone or traditional RAG embeddings. Why? Because an individual's entire career history even a 20-year veteran's is only a few kilobytes of text. It easily fits into a modern LLM's context window. Instead of vector search, we use In-Context Semantic Routing . We feed your entire JSON database to the LLM and prompt it to output an array of bullet ids that semantically match the Job Description. The LLM acts as the retriever, but the actual text insertion is handled deterministically by Python. Is it really "Zero-Trust" if the AI generates the Cover Letter and Profile? Yes, but it's a Constrained Generation Boundary . While your hard facts experience bullets, projects are 100 % immutable and fetched via IDs, your Profile and Cover Letter must be dynamically written to address the specific company's mission. To maintain Zero-Trust here, the Python compiler cv compiler.py intercepts these generated text fields before rendering the LaTeX. When the LLM analyzes the Job Description, it is forced to populate a missing skills array for any required skills you do not possess. The compiler performs a strict Regex negative lookahead/lookbehind intersection ?< \w between your missing skills list and the AI-generated free-text. If a match is found, the compiler immediately throws an EigenTruthViolationError and aborts. The AI cannot sneak missing skills into your prose to trick the ATS scanner. Why a Regex engine and not a State-of-the-Art "LLM-as-a-Judge"? Because Pragmatism Over-engineering . Your CV compiler should run completely offline, deterministically, in <1 second, without requiring an active API key or costing $0.05 per build. Our EigenTruth Engine is O 1 , free, and bulletproof for this specific use case. LLMs are notoriously bad at outputting exact string matches. If your database ID is aws migration 2023 , the LLM might output aws mig 23 . Traditional build systems would throw a KeyError and crash. EigenCV uses the rapidfuzz library C++ optimized Levenshtein distance to mathematically calculate the closest matching ID in your database. This provides graceful degradation: we tolerate minor AI typos while strictly maintaining data integrity. We do not rely on complex "Prompt Engineering" to format the CV. We force the LLM to output data strictly matching a cv schema.py Pydantic model. If the LLM violates the schema e.g., nesting arrays where strings should be , the pipeline rejects it before it ever reaches the LaTeX compiler. Strong typing beats prompt engineering. Many modern CV builders use web technologies React/HTML and use Headless Chrome Puppeteer to print to PDF. This is fatal for ATS. Browser engines often render PDFs with fragmented, non-sequential text layers, causing ATS parsers to misread the document. EigenCV uses pdflatex . LaTeX was built from the ground up for print typography. It guarantees a sequential, machine-readable text layer, ensuring the ATS parses your resume exactly from top to bottom. Looking to customize the LaTeX templates, add your own personal dossier for cultural-fit Cover Letters, or understand the Pydantic schema? Most commercial AI resume builders optimize for feeling good, not for technical accuracy. By keeping your resume in a database and restricting the AI to basic data retrieval, you keep 100 % control over your facts . You only automate the tedious parts: LaTeX formatting and ATS keyword matching. Your career is a database. Version control it. Stop maintaining 15 different Word documents. By keeping your career facts in JSON files, you can treat your resume like software. Fork this repo, make it private, and use Git to track your career progression git commit -m "Promoted to Senior" . When you find a job you like, let the Agent query your database, compile your LaTeX, and land the job. EigenCV's default LaTeX templates utilize a color palette inspired by the excellent Awesome-CV https://github.com/posquit0/Awesome-CV project created by Byungjin Park posquit0 . We are grateful for their beautiful typographic colors Note: The structural layout of EigenCV is entirely custom-built for Jinja2 deterministic rendering and was not derived from Awesome-CV's layout engine .