📌 Context & Background

This project was developed as an automated demonstration system for a customer support team at a mid-sized service company processing approximately 300–500 support call recordings per month. Before automation, quality assurance (QA) team leads manually listened to a small sample (5–10%) of call recordings to evaluate call quality, agent compliance, and customer sentiment, manually recording notes in Google Sheets. This selective process left over 90% of calls unmonitored, resulted in slow feedback loops for agents, and delayed detection of customer churn risks. (Educational / Demo Project based on realistic business parameters).

📌Tech Stack: n8n, AssemblyAI API, OpenAI API (LLM), Google Sheets, Google Drive.

📌 Problem Statement

• Manual Effort: QA managers spent hours manually listening to full call audio files to evaluate performance and customer sentiment.
• Lack of Data & Visibility: No real-time structured tracking existed for call metrics such as speech duration, word counts, or transcription costs.
• Human Error & High QA Costs: Manual sentiment classification was subjective and inconsistent, while scaling manual call reviews required dedicated full-time staff

📌Solution:

• Two-Part Modular Architecture: Built a main data processing pipeline alongside an autonomous analytics AI Agent.
• Automated Audio Transcription & Sentiment Analysis: Workflow automatically retrieves audio files from Drive/URLs, transcribes them using AssemblyAI with timestamps, and analyzes sentiment, key topics, and concise summaries via OpenAI.
• Metric Calculation & Structured Logging: Automatically computes audio duration, word count, and estimated API transcription cost, logging everything to Google Sheets and n8n Data Tables.
• Autonomous Analytics AI Agent: Implemented an independent chat interface equipped with a custom tool (Call Tool) to query processed tables on demand and generate high-level performance trend reports.

📌 Process (Project Lifecycle)

• Discovery & Requirements Gathering: Mapped the QA review workflow, identified key speech-to-text requirements, and defined output parameters (sentiment, summary, topic, cost).
• Architecture Design: Designed a two-part modular system: a deterministic ETL pipeline for call ingestion/processing and a conversational AI agent for reporting.
• Implementation & Integrations: Configured n8n triggers, integrated AssemblyAI for transcription, developed LLM prompts for sentiment scoring, and set up n8n Data Tables with custom agent tools.
• Testing & Prompt Tuning: Tested the workflow against diverse sample audio recordings (varying lengths, accents, and emotional tones) to verify transcription accuracy and sentiment consistency.
• Deployment: Executed end-to-end test runs in a staging environment to validate audio fetching, API error recovery, and database writes.
• Error Handling & Optimization: Integrated a dedicated Error Workflow to isolate failed API calls, prevent job crashes, and maintain real-time cost tracking logs.

📌 Results & Validation