Support Ticket Analysis & Response Automation

Customers in Pakistan write support tickets in Roman Urdu ("mera order kab aayega"), often code-switched with Urdu script and English in a single message.

Tickets come in by email or web form and get triaged in one structured call: category (8 classes), urgency, sentiment, and key entities (order ID, product). A reply draft is then generated in the customer's own language, grounded in your uploaded knowledge-base policy documents (PDF/DOCX/MD) and checked for groundedness. Nothing ships without a human review: tickets land in a review-first inbox where an operator approves, corrects, or discards each draft.

The engine is a QLoRA fine-tune of Qwen2.5-3B-Instruct that hits 80.4% category accuracy (0-shot: 25%; an unnamed 70B baseline: 63.5%) with 100% valid JSON output, at zero per-inference cost on local hardware via Ollama. One Pydantic schema and one prompt template are shared across training, evaluation, the API, and the UI; no drift between what was trained and what runs.

Published to Hugging Face: LoRA adapter (~57 MB), dataset, GGUF weights, and a live Gradio demo Space that runs the fine-tuned model on ZeroGPU in the browser.