SupportGen — AI Customer Support Ticket Automation Agent

1. Elevator Pitch

Problem: Customer support teams spend significant time manually triaging, summarizing, and drafting responses to repetitive support tickets, leading to slower response times and inconsistent resolutions.

Solution: SupportGen automates the end-to-end support ticket workflow through a coordinated multi-agent system capable of classification, summarization, knowledge-based response drafting, and final quality assurance.

Value: Reduces manual triage and drafting time by 60–80%, improves accuracy and consistency, and allows human support teams to focus on complex, high-impact issues.

---

2. Core Concept & Novelty

Why Agents?

Static ML systems fail at tasks requiring reasoning, multi-step workflows, or tool usage. Agents solve this by enabling:

• Reasoning through ambiguous or incomplete tickets
• Decomposing tasks into multiple steps
• Calling APIs and internal tools
• Maintaining context across the full ticket lifecycle

Key Concepts Demonstrated

• Multi-Agent Pipeline:
  Classifier → Summarizer → Response Generator → QA Validator → Publisher

• Tool Use:
  Mock ticket API, knowledge base search, structured output validation

• Sessions & State:
  Ticket-level context tracking with memory

• Long-Term Memory:
  Stores recurring issue signatures & historical customer patterns

• Observability:
  Structured logs, traces, and evaluation pipeline

---

3. Architecture

SupportGen uses a sequential multi-agent design:

1. Ingestion Agent — retrieves new tickets
2. Classifier Agent — predicts category and priority
3. Summarizer Agent — generates concise TL;DR summaries
4. Responder Agent — drafts responses from knowledge base
5. QA Agent — validates tone, accuracy, and completeness
6. Publisher Agent — sends final response & updates ticket status

Architecture Diagram:
Add your diagram to:

docs/architecture.png

---

4. How to Run (Reproducible)

Prerequisites

• Python 3.10+
• pip

Install Dependencies

pip install -r requirements.txt

Start the Mock Ticket API

python src/tools/mock_ticket_api.py

Run the Demo

python src/agent.py --demo

Run Evaluation

python src/eval.py --eval-set data/eval_set.csv

---

5. Files & Structure

supportgen-capstone/
├── README.md
├── report.pdf
├── requirements.txt
├── src/
│   ├── agent.py
│   ├── llm_client.py
│   ├── memory.py
│   ├── logger.py
│   ├── classifiers.py
│   ├── generator.py
│   ├── eval.py
│   └── tools/
│       ├── ticket_adapter.py
│       └── mock_ticket_api.py
├── notebooks/
│   └── demo.ipynb
├── data/
│   ├── sample_complaints.json
│   └── eval_set.csv
├── docs/
│   ├── architecture.png
│   └── metrics.png
├── deploy/
│   ├── Dockerfile
│   └── cloud_run.md
├── tests/
│   └── test_agent.py
├── logs/
│   └── example_log.jsonl
└── scripts/
    └── evaluate.sh

---

6. Evaluation

Dataset

Synthetic customer ticket dataset including:

• Refund requests
• Bug reports
• Account access issues
• Billing questions
• General inquiries

Metrics

Metric	Description	Result
Classification Accuracy	Correct ticket category prediction	TBD
ROUGE-L	Summary similarity score	TBD
Response Quality	Human-rated usefulness	TBD

Full evaluation report available at:

reports/eval_report.json

---

7. Deployment

Option A — Docker

docker build -t supportgen .
docker run -p 8080:8080 supportgen

Option B — Cloud Run

1. Build and push container
2. Deploy using gcloud run deploy
3. Expose /ticket endpoint

---

8. Video

YouTube link:

Your video should cover:

• Problem overview
• Why agents?
• Architecture explanation
• Live demo
• Evaluation summary
• Deployment walkthrough

---

9. Limitations & Future Work

• Multi-intent tickets may still require human review
• Knowledge base content must be kept updated
• Currently supports only English
• Future improvements:
  • RAG-based retrieval
  • Automated escalation rules
  • Multi-language support
  • Dashboard for analytics & monitoring

---

10. Team

• Pranav Gaikwad — Development, Agent Architecture, Video Creation
• Sanchali Torpe — Evaluation, Testing, Documentation

11. License & Credits

This project is licensed under the MIT License, allowing open use, modification, and distribution for both academic and commercial purposes.

Credits

• Developed as part of Agents Intensive — Capstone Project 1 (2025)
• Designed using modern multi-agent patterns including tool-calling, memory, observability, and evaluation
• Includes a mock ticket API, synthetic datasets, and evaluation scripts created exclusively for demonstration purposes
• Thanks to the Google Agents Intensive team and Kaggle community for guidance and resources