MII Microsoft AI Portfolio and Assistant

Situation

MII's Microsoft AI practice needed a public portfolio of its non-confidential work, with an assistant visitors could question. The assistant had to answer only from curated content with citations, resist prompt injection and stay within an enforceable daily budget, while the team needed to publish projects and read analytics without an engineer. I was the sole engineer.

Role and key decisions

1. 1

   Put a deterministic router in front of three YAML agents

   Why: The assistant runs on a FastAPI backend with a deterministic router and three YAML agents on Microsoft Agent Framework; when a model path fails, a guarded fallback tier keeps the guardrails in place.

2. 2

   Gated answers on citations and checked output sentence by sentence

   Why: The assistant may answer only from curated content, so Cosmos DB vector search and Azure AI Search hybrid retrieval feed citation gating, with Prompt Shields on the input and sentence-level checks on the output.

3. 3

   Fixed production failures that were not being surfaced

   Why: Broken stream framing, telemetry reporting zero requests and a daily spend cap that undercounted tokens were not being surfaced. I fixed each one.

Results

• -23%

  Reduction in prompt tokens; model calls per turn fell from 3 to 2

  Measured in pull-request benchmarks on a low-traffic site, about 33 turns per day.

• ~4.2x

  Token undercount identified and corrected in the daily spend cap

  Identified during a production review.

Lesson

A RAG assistant earns trust by answering only from curated content and showing where each answer came from. The operational side, such as stream framing, telemetry and the daily spend cap, needs watching too, because failures there can stay silent.

Architecture

Gallery

Tech stack

• Next.js
• FastAPI
• Microsoft Agent Framework
• Cosmos DB
• Azure AI Search
• Azure OpenAI
• Azure AI Content Safety
• Microsoft Entra ID
• App Service