How manishgaurav24/resume_poc works
Project type and purpose
• This is an Agentic AI–based document transformation tool (open-source engine) that accepts arbitrary PDF resumes, extracts structured data, normalizes and maps it to a fixed JSON schema, and finally renders a pixel-perfect Exavalu-style PDF.
• It uses a LangGraph workflow of discrete “agents” (extract, normalize, map, render).
File structure
• Top-level:
• main.py & graph.py – entry point & DAG definition
• agents/ – four single-responsibility Python modules (extract_agent.py, normalize_agent.py, map_agent.py, render_agent.py)
• render/ – template code (exavalu_template.py) + static assets (logo, watermark)
• utils/ – shared helpers (gemini_utils.py, pdf_utils.py, path_utils.py)
• resumes/ – input PDFs
• Raw_Json/ & Modified_Resumes/ – generated JSON and output PDFs
• Configuration: pyproject.toml, .python-version, .gitignore, .env
README for additional insights
• It defines a 4-stage pipeline:
1) EXTRACT_AGENT: hybrid extraction (PyPDF + Gemini API)
2) NORMALIZE_AGENT: cleans & merges raw data
3) MAP_AGENT: dynamic JSON → fixed Exavalu schema (schemas/*.json)
4) RENDER_AGENT: ReportLab PDF generation with layout rules
• It also mentions fallback logic if Gemini quota is exceeded.
a. Main components (use rectangular nodes):
• Input Storage: “resumes/” (PDF files)
• Orchestrator: LangGraph engine (graph.py)
• Agents:
• Extract Agent (extract_agent.py)
• Normalize Agent (normalize_agent.py)
• Agent (map_agent.py)
• Render Agent (render_agent.py)
• Schemas: dynamic_resume.schema.json, exavalu_fixed.schema.json
• Template & Assets: exavalu_template.py + logo/watermark
• Output Storage: Raw_Json/ (JSON) & Modified_Resumes/ (PDF)
• External Service: Google Gemini API
resumes/ → (PDF stream) → EXTRACT_AGENT → produces raw JSON → Raw_Json/
Raw_Json/ → NORMALIZE_AGENT → normalized JSON → (in-memory)
normalized JSON → MAP_AGENT → fixed-schema JSON
fixed-schema JSON + exavalu_template → RENDER_AGENT → output PDF → Modified_Resumes/
c. Architectural patterns & design principles:
• Agentic pipeline: single-responsibility nodes
• Modular utils for PDF & path operations
• Schema-driven mapping for data validation
• Graceful degradation (Gemini → PyPDF)
d. Technologies & frameworks:
• LangGraph (workflow orchestrator)
• Google Gemini 2.5 Flash (multimodal extraction)
• PyPDF (text extraction fallback)
• ReportLab (PDF rendering)
• Python (.env for API keys, uv for dependency management)
Detail
• Every folder/component (utils, render/assets, schemas)
• LangGraph orchestration layer as a central spine
• Modules (gemini_utils.py, pdf_utils.py) to show where specialized logic resides