hiver-ai-support-agent

This project is an end-to-end AI support agent pipeline designed to parse real, noisy customer support tweets, classify customer intents accurately, retrieve historical ground-truth replies, and use LLMs to draft context-aware responses while safely routing critical issues to human agents. It is built for the Hiver SDE Intern Take-Home Assignment, targeting the @AmazonHelp brand, and uses the Customer Support on Twitter dataset (~3M tweets, Kaggle). The system classifies incoming customer messages, drafts grounded replies, and makes escalation decisions. The pipeline includes four main components: an intent classifier using GPT-4o-mini with few-shot learning, a historical retriever using TF-IDF cosine similarity, a reply generator using GPT-4o-mini, and an escalation router combining rules, heuristics, and LLM. Evaluation is performed on a golden set of 144 hand-labelled examples, with metrics such as intent accuracy, BLEU, ROUGE-L, and escalation F1. The system does not handle real-time conversation management, account-level actions, multi-language support, image/media understanding, or fine-tuned models.