Omprakash JenaEnterprise AI Architect

Global aerospace engine manufacturer

GenAI email search & summarisation

Manufacturing · Aftermarket

Production

Hybrid search and AI summaries over a million historical emails — query time down from about 30 minutes to under 90 seconds.

My role
AI Solution Architect
Delivery
Team of 5
Status
In production

<90 s

per query, down from about 30 minutes

Problem

Engineers lost around 30 minutes per query digging through a million historical emails with slow, inaccurate mailbox search.

Approach

Hybrid retrieval — semantic vectors with BM25 keyword matching — behind LLM guardrails, with AI summaries shown next to the source email's metadata.

Architecture

Deployment view. Engineers on the corporate network sign in with Microsoft Entra ID, which provides SSO and managed identities, and reach Azure Front Door, which serves a React single-page app from a Storage static website. The app calls API Management, which routes requests to FastAPI search and summary microservices on AKS. The services cache access tokens in Azure Cache for Redis, fetch tokens on refresh from an access management service and write logs to Azure Log Analytics. In a Databricks AI platform workspace they run hybrid search on a Lakebase vector and BM25 index, generate summaries through Model Serving, which calls Azure OpenAI LLM and embedding models, and send traces to MLflow; they also rerank results with an Azure AI Foundry reranker. On premises, an email extractor on a Windows host reads Outlook shared mailboxes over MAPI and writes Parquet exports to a NAS landing share. A self-hosted integration runtime, orchestrated by an Azure Data Factory daily pipeline, copies them daily into bronze in a separate Databricks data workspace; bronze is cleaned and conformed into silver and curated into gold, and gold data is embedded and indexed into the Lakebase index.

Outcome

Under 90 seconds per query, released to production with a cleared quality gate and full observability.

Technology stack

  • Azure Front Door
  • Azure Storage
  • API Management
  • AKS
  • Azure Cache for Redis
  • Azure Log Analytics
  • Azure Databricks
  • Databricks Model Serving
  • Databricks Lakebase
  • MLflow
  • Azure OpenAI
  • Azure AI Foundry
  • Azure Data Factory
  • Microsoft Entra ID
  • FastAPI
  • React
  • Python

Client and project names are withheld. Architecture and outcomes are described in general terms to respect client confidentiality.