Skip to main content
BACK_TO_WORKINDUSTRY: MANUFACTURING

Agentic AI & MCP Platform

Built an internal AI platform that gives engineers and applications governed, auditable access to enterprise data through MCP, plus the retrieval and agentic workflows that run on top of it.

PythonFastMCPAnthropic ClaudeAzure AI FoundryPostgrespgvectorSQL Server
0.1 OUTCOME_METRIC
Ad-Hoc AI → Governed Platform, In Daily Use

The Problem

Engineers and internal applications at a multi-plant manufacturer needed real answers from operational and ERP data, but getting them meant one of two bad options: write one-off SQL against a sprawling production schema every time, or paste data into public AI tools with no governance at all.

The AI experiments that did happen were ungoverned. No access control, no audit trail, no shared memory between sessions, and no reliable way to tell whether an AI answer was actually correct. Every new workflow re-solved data access from scratch, and nothing built on what came before. For a business where the data drives production decisions, that combination of no governance and no way to measure quality was a wall.

What I Built

An internal AI platform on Python and FastMCP that gives engineers and internal applications governed, auditable access to enterprise data and AI tooling through the Model Context Protocol (MCP). I built it with Claude Code as my primary development environment throughout. It is in daily use by the engineering team, with production deployment underway.

Governed MCP Tool Surface

The platform exposes SQL Server and operational data as governed MCP tools, so an engineer or an agent gets exactly the data it needs through a controlled, audited interface instead of raw database credentials or copy-paste. Every call is scoped and logged. A persistent context and memory store is scoped across personal, project, team, business-unit, and organizational levels, so knowledge accrues over time instead of evaporating at the end of each session.

Retrieval, Search & Evaluation

Designed the search stack over a Postgres and pgvector knowledge base: hybrid lexical, semantic, and contextual retrieval, with reranking, query rewriting, and metadata filtering. Added image and vision search over SOPs and technical diagrams. Built LLM-as-a-judge evaluation pipelines so retrieval and answer quality are measured rather than assumed, and regressions surface before users hit them.

Agentic Order Entry (Human-in-the-Loop)

Developing an agentic workflow that ingests customer purchase orders spanning hundreds of products and millions of configurations and produces production-ready manufacturing orders for human review before anything commits. It combines the platform's MCP tooling, Azure AI Foundry, model evaluation, and structured extraction. A large part of the work is upstream: driving standardization of engineering data so the AI has reliable inputs, because an agentic workflow is only as trustworthy as the data underneath it.

Governed AI Assistant

Designing a page- and role-aware AI assistant for the company's internal platform, powered by the platform's tool surfaces through Azure AI Foundry, with defense-in-depth access gating and per-user usage limits. The assistant can read and cite the right records for the person asking, and nothing else, by construction.

The Outcome

Replaced ad-hoc, ungoverned AI access with one governed, audited MCP platform in daily use by the engineering team
Gave engineers and internal applications a single, controlled path to operational and ERP data, with every call scoped and logged
Built retrieval and LLM-as-a-judge evaluation so AI answer quality is measured and regressions are caught before users are
Established a human-in-the-loop pattern for agentic workflows: the AI drafts, a person approves, nothing commits unreviewed
Created the shared foundation that the internal AI assistant and the agentic order-entry workflow both build on