tw. ← back to work
// case study
ai enablement

LLM/RAG Translation Engine

Amway · greenfield build, delivered to production · 100+ markets

A manual, weeks-per-asset translation process, rebuilt as a governed LLM/RAG system: from a blank page to production across 100+ markets, with no existing architecture, vendor, or delivery plan to inherit.

100+ markets
in production
$1.3M
projected annual savings
4 parallel
workstreams
governance reused
as company standard

the problem

Amway's 1M+ independent business owners need marketing and product content in their local languages across 100+ markets. The existing process was entirely manual: content created centrally, routed to regional teams, translated by local agencies, a cycle that took weeks per asset and cost millions annually. I was given ownership of the end-to-end delivery of a replacement, a greenfield LLM/RAG build, zero to production.

the approach

The first move was a delivery plan that forced scope clarity before any code was written: what we were building, what we weren't, and what "done" meant. Four workstreams had to move in parallel: engineering (architecture and build), legal (IP review of LLM outputs and data privacy across jurisdictions), vendor (model selection, contracting, SLAs), and 100+ regional stakeholders who needed to validate quality thresholds in their own languages.

I owned the acceptance criteria directly: confidence thresholds below which outputs required human review, escalation criteria for edge cases like legal-sensitive or brand-specific terminology, and the human-in-the-loop checkpoints that gave regional teams enough control to actually trust the system. Where legal and regional stakeholders disagreed, mainly on which content categories needed mandatory human review, I mediated the tradeoff, documented it, and got executive sign-off rather than letting it stall delivery.

the outcome

Delivered to production across 100+ markets. Projected $1.3M in annual savings from reduced agency translation costs. The governance model built for this system, confidence thresholds, human-in-the-loop criteria, escalation paths, became the operational standard applied to subsequent AI deployments at Amway, including the broader AI transformation program.