How bertughas123/normvision works
Project type and purpose
• NormVision is a data-processing pipeline rather than a traditional full-stack web app. It ingests unstructured PDF visit reports, applies NLP/ML/AI to extract and normalize fields, computes KPIs, then outputs structured reports and synchronizes with a CRM.
File structure
• Top-level “extractor/” modules handle all stages of raw PDF → text → normalized data:
• pdf_reader.py → notlar_parser.py → sections.py → normalize.py → schema.py → llm_fill.py → campaigns.py
• “analyzer/” modules compute financial and sales KPIs.
• “bridge/” layer pushes results into a CRM (sales_visit_bridge.py).
• Orchestrator scripts (runner_step1.py, runner_batch.py, runner_weekly.py, runner_monthly.py) coordinate flows for single, bulk, weekly, and monthly runs.
• Outputs land in Reports/ as CSV/Markdown.
• External dependencies: Google Gemini AI (via llm_fill), PyMuPDF/pdfplumber/pytesseract for PDF/OCR.
Extract insights from the README
• Key feature areas map cleanly to modules: extraction, AI-powered completion, KPI analysis, campaign tracking, report export, CRM sync.
• The README emphasizes a modular pipeline that can be run in different modes (step1, batch, weekly, monthly).
Main components
a. Input Layer
• PDF Source (single or directory)
b. Extraction Layer
• PDF Reader
• Parser & Section Splitter
• Normalizer & Schema Builder
• AI Completion (Gemini LLM)
• Campaign Extractor
c. Analysis Layer
• Financial Analysis Module
• Sales Performance Module
d. Integration Layer
• CRM Bridge (SalesVisitBridge)
e. Orchestration Layer
• Runners (step1, batch, weekly, monthly)
f. Output Layer
• Report Generator (CSV, Markdown)
• Reports/ Directory
g. External Services
• Google Gemini AI
• CRM API
Relationships and data flow
• Runners invoke the Extraction Layer in the correct order.
• Extracted & normalized data flows into Analysis Layer.
• Analysis results either feed into the Report Generator or into the CRM Bridge.
• Reports are written to disk; CRM Bridge calls external CRM API.
Architectural patterns and principles
• Layered pipeline: clear separation of concerns (extract → analyze → publish).
• Modular design: each module does one job and can be tested independently.
• Orchestration vs. step-function: runners encapsulate scheduling logic.
Relevant technologies to annotate
• Python 3.9+ (core language)
• PyMuPDF/pdfplumber/pytesseract (PDF & OCR)
• Google Generative AI (Gemini API)
• pandas / CSV / Markdown for outputs
• CRM SDK (used in bridge)