s a Foundry AI Developer, you will be at the forefront of building the “Enterprise Brain” by architecting intelligent systems that bridge the gap between massive data lakes and actionable AI. Your primary objective is to design, develop, and optimize an agentic ecosystem using Microsoft Fabric, Azure AI Search, and the Model Context Protocol (MCP). You will be responsible for creating the “nervous system” that allows LLMs to interact securely with proprietary data and internal tools while maintaining peak performance and cost efficiency.

Key Responsibilities

1. Unified Knowledge Engineering (Fabric & AI Search)
The Focus: Building the "Enterprise Brain" by linking Microsoft Fabric’s OneLake to Azure AI Search / Foundry IQ.
Key Responsibilities:
Developing Fabric Data Agents that serve as live conduits between structured lakehouse data and the Foundry Hub.
Configuring Azure AI Search as a vector-enabled index to ingest unstructured data from Fabric, enabling "Agentic RAG" where the agent decides when to pull deep historical data vs. real-time telemetry.
Implementing Foundry IQ skills to handle semantic chunking and multimodal embeddings, ensuring the LLM receives highly relevant context.
2. Python-Based MCP Server Architecture
The Focus: Creating a secure "Nervous System" that allows LLMs to act on proprietary systems using the Model Context Protocol (MCP).
Key Responsibilities:
Developing custom MCP Servers in Python (using asyncio and FastAPI) to expose private APIs and local tools as "Functions" that the agent can invoke.
Hub Registration: Registering these servers within the Foundry Hub as persistent connections, allowing multiple AI agents to share the same toolsets securely.
Managing identity and access through Foundry Control Plane, ensuring that agent "tool-use" follows enterprise security policies (OAuth/Managed Identities).
3. Model Orchestration & Token Optimization
The Focus: Selecting the "Right Model for the Right Task" to maximize performance while minimizing token burn.
Key Responsibilities:
LLM Evaluation: Using Foundry’s Evaluation Dashboard to run "LLM-as-a-judge" benchmarks, comparing models (e.g., GPT-4o for complex reasoning vs. Phi-3 or Llama 3 for extraction) to find the most cost-effective fit for specific sub-tasks.
Token Efficiency: Designing Python-based "Routing Logic" that sends simple queries to small language models (SLMs) and reserves expensive LLMs for complex orchestration.
Observability: Monitoring "Token Burn" per agent run in the Foundry portal to identify prompt bloat and optimize system instructions for brevity without losing accuracy.