Smart Reply Copilot: Multi-Step Prompt Chaining

What I built

Customer support agents spent ~10 minutes manually drafting each ticket response, with inconsistent tone and classification quality across the team.

Build an AI copilot that automates ticket triage and response drafting while maintaining human oversight and measurable accuracy at each stage.

What I did

Engineered a 3-step prompt chain (category detection → tone calibration → draft generation) with confidence scoring at every step. Built a manual quality scoring system validated against 22 real support tickets. Deployed to production on Render with environment-based configuration and modular architecture for independent step optimization.

What came out of it

Achieved 95.5% category classification accuracy and 86.4% tone calibration accuracy. Average confidence score of 0.92 across the full chain. Reduced response time from ~10 minutes to ~2 minutes per ticket—an 80% reduction.