TelecomIQ — Telecom Complaint Intelligence & Automated Resolution Assistant

---

📌 Overview

TelecomIQ is an AI-powered complaint intelligence platform for telecom operators. It automatically classifies complaints, detects customer sentiment, predicts escalation risks, recommends resolutions, and generates ticket summaries — orchestrated end-to-end using a LangGraph agentic pipeline.

Use Case: Telecom Complaint Intelligence & Automated Resolution Assistant
Dataset: Kaggle — ravillatejakumar/telecom-complaints-monitoring-system
Type: NLP / Information Extraction

---

👥 Team

#	Member	Ownership	Main Responsibility
1	Abhyanshu	Business + Dataset	Problem definition, dataset, EDA, business impact
2	Srishti	NLP Metadata	Keywords, NER, speaker identification, time segmentation
3	Yashraj	ML + Sentiment	Classification (TF-IDF + Logistic Regression), VADER sentiment
4	Vibhuti	RAG	Retrieval, vector similarity, knowledge base
5	Vaibhav Raj	GenAI	LLM prompts, resolution generation, ticket summary
6	Veer	Risk + Compliance	Priority scoring, escalation prediction, PII detection, human-in-loop
7	Vishant	Full Stack + Architecture	React, FastAPI, database, integration, deployment

---

🔄 End-to-End Pipeline

Customer Complaint
       ↓
  Input Validation
       ↓
  Text Classification        (TF-IDF + Logistic Regression — 89.12% accuracy)
       ↓
  Sentiment Analysis         (VADER + TextBlob)
       ↓
  Priority + Escalation Risk (Multi-factor scoring → CRITICAL / HIGH / MEDIUM / LOW)
       ↓
  Vector Historical Search   (Cosine similarity over 2,200+ indexed tickets)
       ↓
  RAG Knowledge Base         (TF-IDF over 11 telecom SOP documents)
       ↓
  LangGraph Orchestration    (7-node StateGraph)
       ↓
  GenAI Triage Assistant     (Groq Llama-3.3/Qwen → SOP fallback)
       ↓
  Resolution + Ticket Summary
       ↓
  Support Agent / Dashboard

---

📂 Dataset

Property	Value
Source	Kaggle — ravillatejakumar/telecom-complaints-monitoring-system
Raw records	2,224
After deduplication	2,204
Train / Val / Test split	70% / 15% / 15%
Test accuracy	89.12%
Weighted F1	0.8903

---

🧪 ML Validation Results (Held-Out Test Set — 331 Samples)

Category	Precision	Recall	F1-Score	Support
Billing Dispute	0.987	0.802	0.885	96
Broadband Performance	0.902	0.974	0.937	38
Call Drops	0.667	1.000	0.800	4
Cancellation	0.750	1.000	0.857	3
Customer Service	0.889	0.889	0.889	9
Data / Usage Issue	0.944	0.971	0.958	35
Equipment / Router	0.500	1.000	0.667	1
Installation	1.000	0.333	0.500	3
Network Connectivity	0.000	0.000	0.000	1
Service Outage	0.539	0.778	0.636	9
Service Request	0.872	0.932	0.901	132
Overall	0.902	0.891	0.890	331

---

🛠️ Technical Stack

Layer	Technology
Agentic Orchestration	LangGraph StateGraph (7 nodes)
ML Classification	Scikit-learn TF-IDF + Logistic Regression
Deep Learning	DistilBERT (offline fallback), BART zero-shot
Sentiment Analysis	VADER + TextBlob
GenAI	Groq (Llama-3.3 / Qwen) → SOP fallback
Vector DB / RAG	TF-IDF cosine similarity over 2,200+ complaints + 11 SOP docs
Backend	FastAPI + SQLAlchemy + SQLite/PostgreSQL
Frontend	React 19 + Vite
Deployment	Vercel (frontend + backend)

---

🚀 Local Setup

Backend

cd backend
pip install -r requirements.txt
python scripts/train_kaggle_dataset.py   # download dataset, train models, seed DB
python start_backend.py                  # runs on http://localhost:8000

Frontend

cd frontend
npm install
npm run dev                              # runs on http://localhost:5173

Demo Access

• Any email containing admin → Admin Dashboard
• Any email containing agent → Agent Queue
• Any other email → Subscriber view

---

📄 License

MIT License — built for the Cognizant NPN AI & Analytics evaluation.