Enterprise Generative AI Platform (Case Study)

Cover Image for Enterprise Generative AI Platform (Case Study)

A company-wide generative AI application rolled out to all employees of a major Japanese convenience-store retailer. I serve as technical lead and architect at Accenture Song Japan, leading full-stack delivery across the LLM layer, backend, agents, and UI. The application reached over 80% answer accuracy in production.

Client details, screenshots, and commercial figures are withheld for confidentiality. The diagram above is a simplified, generic view of the architecture.

My role

Technical lead and architect, responsible for the architecture and for hands-on delivery across the stack, from the LLM integration layer to the user interface.

What I designed and built

  • Multi-model LLM architecture. A single platform that integrates Anthropic, OpenAI, and Google models. I prototyped on GCP Vertex AI and moved production to Azure to meet the client's security requirements.
  • Agent layer. Tool calling and external system integrations, with permission-based authorization that controls what each user and each agent can access.
  • LLM observability and evaluation. Langfuse for tracing, prompt versioning, quality monitoring, and cost tracking, so answer quality and spend can be measured and improved over time.
  • Business case. I authored the proposal quantifying the multi-year cost savings the client can expect from AI-driven automation.

Stack

Python, Django, TypeScript, Vue.js, Nuxt.js, Node.js, Azure, GCP Vertex AI, Langfuse, Anthropic / OpenAI / Google APIs