Telecom traffic analysis system with AI-Vision
Objective: Automate the critical business process of a telecom company — analyzing incoming rate sheets from VoIP traffic suppliers. The problem was the variety of formats: suppliers send prices in Excel, CSV, PDF, and even images (screenshots in messengers). Manual processing took hours, leading to missed profitable deals in a dynamic market. Requirement: The system must be "omnivorous," identifying profitable (BUY) and unprofitable (SELL) directions by comparing them with the internal market API and instantly notifying managers.
My Contribution / Solution: The solution was implemented on Self-hosted n8n, using Google Drive, OpenAI (GPT-4o / Turbo), and Google Sheets. The architecture is built on a "Parent-Child" principle for scalability and fault tolerance.
1. Workflow "Omni-Channel Ingestion" (Parent Processes):
Routing and Queue: Implemented Smart Queue logic. The system scans Google Drive, identifies the file type (.xlsx, .csv, .pdf, .png), and processes them one by one (Batch Size: 1) with a 20-minute interval to avoid API overload and limits.
AI Vision & OCR: A cascade was created for processing "unreadable" formats (PDF/Images):
CloudConvert: Conversion of multi-page PDFs into high-quality PNGs (300 DPI).
GPT-4o Vision: Utilization of a multimodal model for visual reading of tables from images, where regular parsers are powerless.
Smart CSV Parsing: For large text files, a "Chunking" algorithm was implemented — breaking the text into packets of 40 lines for processing by a lighter GPT-4 Turbo model, significantly saving the client's budget.
2. Workflow "Analytical Core" (Child Workflow):
Data Enrichment: A complex matching algorithm (JavaScript) was implemented. The system normalizes country names (for example, correcting "Dr Congo" to the official name), determines MCC/MNC codes from the internal directory, and classifies traffic type (Direct/HQ/SS7/Sim) based on file metadata.
Market Intelligence: Integration with an external API (interconnect.solutions). Each line of the price is checked in real-time for market median.
Logic Engine: Automatic margin calculation. The system assigns a BUY status (if the price is below the market) or SELL, and sorts offers from the most profitable.
3. Reliability and UX:
Error Handling: A global error interceptor and local Retry strategies (3 attempts) for unstable HTTP requests were configured.
Reporting: The final result is formed as an interactive HTML report in Telegram with links to the original and processed file, as well as the Top-15 recommendations for managers.
Result: The client received a fully automated traffic procurement department:
Versatility: The system processes any incoming file, from Excel to screen photos.
Response Speed: The time from receiving the file to decision-making was reduced from hours to minutes.
Economic Effect: Managers receive ready "signals" (Buy Alerts) and do not waste time manually comparing thousands of lines.
Stability: Thanks to queues and request optimization, the system operates 24/7 without API failures.
#n8n #OpenAI #GPT4o #ComputerVision #Automation #Telecommunications #VoIP #JavaScript #GoogleDriveAPI #DataEngineering #CloudConvert #TelegramBot
My Contribution / Solution: The solution was implemented on Self-hosted n8n, using Google Drive, OpenAI (GPT-4o / Turbo), and Google Sheets. The architecture is built on a "Parent-Child" principle for scalability and fault tolerance.
1. Workflow "Omni-Channel Ingestion" (Parent Processes):
Routing and Queue: Implemented Smart Queue logic. The system scans Google Drive, identifies the file type (.xlsx, .csv, .pdf, .png), and processes them one by one (Batch Size: 1) with a 20-minute interval to avoid API overload and limits.
AI Vision & OCR: A cascade was created for processing "unreadable" formats (PDF/Images):
CloudConvert: Conversion of multi-page PDFs into high-quality PNGs (300 DPI).
GPT-4o Vision: Utilization of a multimodal model for visual reading of tables from images, where regular parsers are powerless.
Smart CSV Parsing: For large text files, a "Chunking" algorithm was implemented — breaking the text into packets of 40 lines for processing by a lighter GPT-4 Turbo model, significantly saving the client's budget.
2. Workflow "Analytical Core" (Child Workflow):
Data Enrichment: A complex matching algorithm (JavaScript) was implemented. The system normalizes country names (for example, correcting "Dr Congo" to the official name), determines MCC/MNC codes from the internal directory, and classifies traffic type (Direct/HQ/SS7/Sim) based on file metadata.
Market Intelligence: Integration with an external API (interconnect.solutions). Each line of the price is checked in real-time for market median.
Logic Engine: Automatic margin calculation. The system assigns a BUY status (if the price is below the market) or SELL, and sorts offers from the most profitable.
3. Reliability and UX:
Error Handling: A global error interceptor and local Retry strategies (3 attempts) for unstable HTTP requests were configured.
Reporting: The final result is formed as an interactive HTML report in Telegram with links to the original and processed file, as well as the Top-15 recommendations for managers.
Result: The client received a fully automated traffic procurement department:
Versatility: The system processes any incoming file, from Excel to screen photos.
Response Speed: The time from receiving the file to decision-making was reduced from hours to minutes.
Economic Effect: Managers receive ready "signals" (Buy Alerts) and do not waste time manually comparing thousands of lines.
Stability: Thanks to queues and request optimization, the system operates 24/7 without API failures.
#n8n #OpenAI #GPT4o #ComputerVision #Automation #Telecommunications #VoIP #JavaScript #GoogleDriveAPI #DataEngineering #CloudConvert #TelegramBot
Kyiv 
Available for hire
Offer work