Skip to content
← All workIn active development

AI-Assisted PMO Platform

Architected an AI-assisted PMO platform unifying the opportunity pipeline, project delivery and a document knowledge base, with natural-language-to-Cosmos-SQL assistants, streamed SOW generation and a queue-driven ingestion pipeline on Azure Functions.

Role
Solution design · AI architecture & Azure footprint
Year
2026
Type
Full-stack platforms
Stack
Azure OpenAI · Azure AI Search · Cosmos DB · Python · FastAPI · Azure Functions · React · Document Intelligence · Microsoft Graph · Entra ID · Azure App Service

The problem

An IT services company's PMO data was spread across CRM exports, spreadsheets, SOW documents and weekly status decks, so simple questions such as active projects by region or win rate meant opening several files. Status reports and SOWs were assembled by hand, with no single, versioned source of truth.

Architecture

  1. 01

    Ingest

    Uploads land on Azure Storage queues; Node.js Functions run Document Intelligence layout-to-Markdown and versioned Excel diffing.

  2. 02

    Index

    Markdown split into sentence-aware 1,000-character chunks with 100-character overlap, embedded and indexed in Azure AI Search.

  3. 03

    Notify

    Functions post to a notification queue, and FastAPI relays completion events to the browser over WebSockets.

  4. 04

    Route

    Each chat question is classified as opportunity/project data, project core or document knowledge before answering.

  5. 05

    Answer

    Schema-aware prompts generate Cosmos DB SQL whose trimmed results the model summarises, or hybrid search grounds document answers.

  6. 06

    Report

    SOWs stream over SSE into Word; PowerPoint and PDF status reports are emailed through Microsoft Graph.

Engineering highlights

  • 01

    Natural language to Cosmos SQL

    Schema-aware prompts make the model write Cosmos DB SQL under explicit rules: COUNT versus list semantics, latest non-deleted version filters and fixed definitions of active and completed, before results are trimmed and summarised.

  • 02

    Event-driven document pipeline

    Queue-triggered Functions decouple ingestion from the API: layout extraction, sentence-aware chunking, embedding, indexing and blob metadata tagging, with completion pushed to the UI over WebSockets.

  • 03

    Versioned opportunity records

    The Excel processor diffs each row against the latest stored version, appending new records, writing timestamped versions on change and ignoring unchanged rows, so historical pipeline states stay queryable.

  • 04

    Rule-based health scoring

    Project health is computed from RAID severity and escalations, overdue or imminent milestones and open dependencies; the dominant category sets the status and the percentage split drives a health chart.

  • 05

    Dual authentication

    Entra ID tokens validated against JWKS sit alongside internal session JWTs and bcrypt local accounts with refresh tokens, with rate-limited sensitive routes and background expiry of sessions.

  • 06

    Azure deployment design

    Shaped the deployment design and prepared the Azure production cost estimate for an App Service, consumption-plan Functions and Storage-queue topology backed by Cosmos DB, Azure AI Search and Azure OpenAI.

Next projectWarehouse Pallet & Fleet Tracking