8Returns: Building an AI-Native Returns Engine

At 8Returns, I led the transition from a deterministic returns flow (V3) to a fully AI-native system (V4), redesigning how merchants handle customer return requests.

---

🤖 V4 AI Chatbot Architecture

The previous version relied on rigid decision trees and checkbox-based flows.

I designed and implemented V4, a structured LLM-based system powered by the Vercel AI SDK and multiple model pipelines.

AI System Design

The architecture includes three specialized LLM pipelines:

• Structured Return Decision Model
  Produces schema-enforced outputs determining whether the customer should:

  • Refund
  • Exchange
  • Receive store credit
  • Request re-shipment

• Summary Model
  Generates merchant-readable summaries of return conversations.

• Return Risk Matching Model
  Evaluates return patterns and flags risk scenarios.

---

🧪 Evaluation & Observability

• Built prompt evaluation pipelines.
• Created datasets for regression testing.
• Implemented structured prompt testing workflows.
• Monitored model performance, cost, and latency.
• Optimized LLM cost efficiency through model experimentation and routing strategies.

---

⚙️ CI & Infrastructure Improvements

• Implemented automated Neon branch cleanup using the Neon API, preventing environment clutter from merged PR branches.
• Improved CI workflows and development environment hygiene.
• Contributed to frontend architecture alongside a Ruby on Rails backend team.

---

🛒 Shopify-Focused Frontend Engineering

• Built dynamic merchant dashboards.
• Designed AI interaction flows inside Shopify admin environments.
• Collaborated closely with backend engineers to ensure schema consistency and API reliability.
• Delivered seamless UX for AI-driven decision flows.

---

📊 Impact

• Successfully replaced deterministic returns logic with AI-driven automation.
• Reduced manual review processes through structured AI decision outputs.
• Improved cost efficiency of LLM usage via experimentation and routing strategies.
• Established a scalable AI testing and evaluation framework for future iterations.