Devy

Search Posts

Search blog posts by title, description, tags, or content.

Back to About

AI Diagnostic and Feedback Pipeline

Day1 Company · Oct 2025 - Jan 2026

Spring AIGeminiAmazon BedrockOpenAISpring Boot

Tasks

  • Redesigned a diagnosis process that depended on direct HTTP calls and one synchronous prompt into a 7-step Spring AI pipeline: STT load, semantic chunking, STT correction, metric calculation, LLM feedback, question generation, and notification
    • Reduced CS tickets by 98% through better diagnostic quality
    • Integrated the pipeline into the existing Java/Spring environment to reduce maintenance burden and development complexity
    • Separated pipeline stages asynchronously to reduce bottlenecks and stabilize processing time
  • Assigned Gemini, Amazon Bedrock, and OpenAI models to stage-specific roles and designed a back-office flow for changing models and prompts without downtime
    • Reduced LLM cost by about 50%
  • Added hallucination correction and retry logic for format validation failures
    • Reduced diagnostic report generation time from 6-7 minutes to 1-2 minutes

Related Posts