Belle SPA Inc – Internal AI Tools & Automation

GitHub Repo: Junyu06/LLM-Translator

Demo

Problem

Cloud-based SaaS tools for internal workflows incur recurring API costs and raise data privacy concerns for sensitive business operations. Non-technical teams need support for long-form customer inquiries and internal workflows.

Approach

• Designed and deployed an internal LLM system for long-form inquiries and workflow automation used by non-technical teams
• Implemented schema-constrained generation (Pydantic) with explicit validation boundaries for reliable structured outputs
• Deployed local, on-prem inference on idle hardware to reduce recurring API costs and keep sensitive data off external SaaS tools
• Owned end-to-end delivery (design through deployment) with an emphasis on long-context handling and failure recovery

Results

• Reduced recurring SaaS/API spend by shifting inference on-prem
• Kept sensitive business data within internal infrastructure by avoiding external SaaS processing
• Delivered predictable, structured outputs for daily internal workflows with failure recovery

Tech Stack

Python, Pydantic, local LLM deployment, Docker, fault-tolerant pipelines, on-prem inference