Production AI Engineering · Project 04

Ecommerce Refunds on Assistant

First end-to-end conversational refund lifecycle — from intent detection to order resolution — without a live agent.

VP Award

Business recognition for
technical quality & ROI

Contact Deflection

Significant reduction in
live agent contacts

End-to-End

Fully automated refund
lifecycle via conversation

The Challenge

The Problem

Adityagen.ai delivered a production-grade conversational refund system that handled millions of requests through natural language — earning VP recognition for technical quality and measurable contact deflection.

Standard ecommerce refund flows required live agent involvement for most cases — a high-volume, low-complexity workflow that consumed disproportionate support capacity. The opportunity was clear, but the execution was not: building a conversational refund system that felt product-quality (not chatbot-quality) required solving NLU accuracy across real-world intent diversity, managing cross-device session consistency, and integrating directly into the Order Management System without introducing latency that would break conversational flow.

The technical bar was enterprise-grade. This wasn't a prototype — it needed to handle millions of refund requests with the same reliability standards as any production payment system, with escalation paths for edge cases that didn't leave users stranded.

"The measure of success wasn't automation rate — it was whether users could complete a refund through natural conversation without knowing they were talking to a machine."

System Architecture

System Design

Component 01

Custom NLU Model

Intent classification engine trained on large annotated dataset specific to refund interactions. Designed to handle intent diversity across real-world, open-ended user phrasing.

Component 02

Dialog Management System

Hybrid rule + ML dialog management with structured flow control. Handles both straightforward and edge-case refund scenarios with escalation routing.

Component 03

Ecommerce Integration Layer

RPC calls to Order Management System for real-time order and refund data. Dedicated refund source bucket with image retrieval integration. Network-aware refund accountability tracking for auditability and fraud resistance.

Component 04

Native Assistant Integration

Dynamic Composite Frames, Schema Frames, and Account Switcher Frame for native UX. Seamless cross-device and cross-account support.

Key Innovations

Fully automated conversational refund lifecycle — first end-to-end self-service refund flow via Assistant

Native Assistant-grade integration delivering seamless, product-quality user experience

Network-aware refund accountability tracking embedded directly into conversational flow — beyond standard chatbot capabilities

Outcomes & Impact

Results

VP Award Winner

Recognized for technical quality and measurable business ROI

★

Contact Deflection

Significant reduction in live agent workload

Self-Service Adoption

Substantially improved — zero wait time, no escalation

The system achieved significant contact deflection and meaningful operational cost reduction by automating a high-volume, previously agent-dependent workflow. Self-service adoption improved substantially as users could resolve refund requests through natural conversation without wait times or escalation.

The business recognized this impact with a VP Award — reflecting both the technical sophistication and the ROI delivered.

My Role

Cross-Team Leadership & End-to-End Ownership

Spearheaded cross-team collaboration across NLP, ecommerce, and assistant platform teams
Owned end-to-end conversational flow design from intent modeling through OMS integration
Delivered production-grade integration meeting enterprise reliability and performance standards

Tech Stack