Got the Technical Interview — So I Built a Live Demo That Taught Me the Job and Proved I Could Do It

So things moved pretty quick this week with an interview at a local Long Island medical supply company. I'm an AI/ML engineer and a recruiter reached out to me about a role — really a dream job I've been working toward for over three years. I moved through the HR interview quick and decided to build a live demo for the technical interview using FastAPI, Uvicorn, LangGraph, Tailwind, and Claude. Live link, matched the icon and styling to the company's branding. I wanted to be able to share that link and have it come up clean in the Microsoft Teams chat.

Here's the story.

Five years ago I was a commercial fisherman. I loved everything about it — the ocean, the crew, the physical grind, being offshore and in nature. I would have done it forever. Then I broke my neck in a diving accident and lost my ability to provide for myself. I ended up in a nursing home where I spent years learning how to build things — and actually building them to pay for my life outside of the nursing home.

So now that I'm out I'm looking for jobs. A recruiter reached out last week and things moved fast — they want to get their enterprise AI stuff going. I was nervous. I'm visibly disabled and while we were meeting on Zoom it's still obvious. People see the wheelchair before they see the work. So I wanted to lead with what I built first.

It went really well. I weaved in my story about how I built a couple applications, moved out of the nursing home, and actually had the logistics company I built an app for move me into my new place. But anyway — I got to the technical interview.

While I was prepping I realized the best way to learn about the company was to build something with their data. They run Dynamics 365 as their ERP and their API responses use OData which is just JSON endpoints. I'm like — I speak this language. I'm a nerd for exactly this. So what if I build an AI agent that queries their actual product data, has all their technical product information, and I can study while I build? I built tools for checking which parts are compatible with each other, which lots are expiring, and made the human-in-the-loop approval workflow they'd actually need obvious and visible.

That's what I did. In a day.

FastAPI backend with async SSE streaming so you can watch the agent think in real time. LangGraph ReAct agent powered by Claude with 7 tools — 6 read-only data tools and 1 that creates approval items. SQLite database seeded with real product data from their website. API responses formatted as OData JSON to match what D365 actually returns. Single-page app with a chat panel, a live tool activity feed so you can see which data sources the agent is hitting, and an approval queue where every AI recommendation lands for a human to approve or reject.

Key architectural decision: the agent cannot physically modify any data unless approved. There are no write tools for products, inventory, or customers. It can only create approval items. That's enforced in the code, not policy. In a regulated medical device environment where parts end up inside human bodies, that distinction matters.

I also built a data explorer so you can browse the catalog — it's confusing stuff and you want to be able to visualize it. Inventory lots, customer order history, and a compatibility map showing which products connect to each other.

The recruiter told me to walk through my flagship project first. I couldn't do it. I was too excited. I had to show them what I built. When I pulled it up on the video call the live link showed up with the logo — everything matched. The director told me he hasn't seen somebody put that amount of effort into learning about a company in a very long time. He also liked some of the specific tools I built. He said the compatibility mapping — the tool that maps which products connect to each other — was something they've been thinking about for a while.

I don't know if I'm getting the job. I find out Monday. But what I do know is I've been working toward jobs like this for over three years. Every day, all day. And in that journey I've already had a tremendous amount of results — moving out of the nursing home, financing my accessible bathroom, building multiple production apps. I even learned Spanish while I was learning to code.

The demo is still live. I've been checking the server logs and it looks like people at the company might be poking around — it's hard to tell with the SSE connections spamming the network logs, but either way.

I do know eventually I'm going to get that job. I'll put this amount of effort into it until I get what I need. Until I get what I want.