AI-103 · In Progress

Intelligent Insights Platform

A Document Intelligence System built on Azure AI Foundry — OCR and structured extraction via Content Understanding, RAG-grounded summarization and Q&A over a real vector index, responsible-AI evaluation, and a single orchestrator agent that checks its own drafts with an evaluator tool and a computer-vision audit tool. Built while studying for AI-103 (Developing AI Apps and Agents on Azure), the associate-level follow-on to AZ-900 and AZ-104, both already earned.

8Phase 1 milestones complete
Phase 2Currently building
2Certs already earned

This is an honest, in-progress snapshot — not a finished project writeup. The full treatment (architecture diagrams, expandable project detail, security write-up) is drafted and in review, and publishes once Phase 2 is far enough along to go out with it. Until then, this page tracks real milestone status, updated as work happens.

01

What it does

Upload a technical document — a clean PDF export, a flatbed scan, or an angled phone photo — and the pipeline extracts text and structured fields, answers natural-language questions grounded in the document's actual content (not the model's memory), and checks its own output for fabrication before returning it.

Deliberately out of scope: image/video generation and speech modalities — poor fit for a document-intelligence use case. Covered through study material for the exam, not the build. A documented choice, not a gap.

02

Milestones

M0
Complete
Tore down the original West US 3 resource group and all resources — zero carryover into the rebuild.
M1
Complete
Redeployed Azure AI Foundry fresh in West US, with quota re-verified against the new region before building on it.
M2
Complete
Custom document analyzer built and validated against the clean-PDF loan agreement sample document.
M3
Complete
Submit-and-poll extraction pipeline written, debugged, and validated end-to-end against the analyzer.
M4
Complete
Extraction validated across all three image-quality conditions — clean PDF export, flatbed scan, and an angled phone photo of the same document.
M5
Complete
RAG-grounded Q&A working end-to-end against a real Azure AI Search vector index — live-verified with a retrieval-ranking finding: the top-ranked chunk isn't always the correct one, so the pipeline joins every retrieved chunk into context rather than trusting rank alone.
M6
Complete
Evaluator harness built and run twice against independent question sets; selected the production model on confirmed quality parity plus a real ~3x per-token cost advantage, not on assumption.
M7
Complete
Single orchestrator agent over a synthetic content-review scenario — drafts against a template, reuses the M6 evaluator pattern for QA, and adds a computer-vision audit tool covering the one AI-103 exam domain not otherwise touched. Acceptance-tested against a fixed answer key, with its known limits written down rather than smoothed over. Closes out Phase 1 of the IIP build.
M8
Complete
Infrastructure-as-Code baseline opening Phase 2 — Bicep describing the live Foundry account and project, the model deployments at their real capacity, storage, Key Vault and AI Search. Verified with what-if until every remaining difference was a provider-owned property that no template can declare, each one justified in writing rather than deleted to force a clean result.
M9
Not started
Identity foundation — user-assigned managed identities and the role assignments from the access model, in Bicep. The step that makes the keyless migration possible.