I build systems you can open up and understand. Lead backend engineer on a regulated aerospace platform, AI and computer-vision on top. I explain what I build. I rebuild what I want to understand.
No invented metrics. Every number here is one you could check.
The interesting part is never the framework. It's deciding where the truth of a process should live. Before I add AI or automation, I find who's stuck, where the work breaks, and what the system must make explicit. Then I build it so you can open it up and follow it, not take my word for it.
“If the state of a job lives in someone's head, it isn't a system yet.”
Six years building software, most of it backend. Now lead engineer on an AS9110-regulated aircraft-maintenance platform: a .NET monolith behind ten React apps, where the system won't record a sign-off from anyone uncertified to give it. Before any of that, self-taught from a warehouse floor. I lead a team of eight and have mentored thirty-plus interns. Weekends, I rebuild published AI papers from scratch, to understand the models, not just call them.
Try to sign off on a repair you're not certified for, and the platform won't record it: the approval state machine checks your certification before it accepts the signature, one of five types enforced in code across 766 handlers. When a SAP sync half-fails, it stops and flags rather than let the numbers quietly drift.
Five developers rated the GPT-4 output and every rating came back positive, which told me more about my study than about the tool. So I stopped asking people: I swapped two parameters in a function signature, left the name and body untouched, and regenerated. If the docs report the old order, the model is describing what the function is called. They reported the swap. Five for five. It reads the code.
A fault at a retail-outlet terminal used to live in a chat until someone forgot it. Now it's a ticket: pinned to its outlet, owned by a technician, moved New → Process → Finish through the replies on it, and closed after a site visit. A Go API and PostgreSQL model behind a React console.
I reproduce published AI papers from scratch: implement the core method, reproduce the central result, and honestly compare my numbers to the paper's. “Reproduced the trend” and “couldn't reproduce X” are more credible than a number I invented.
A from-scratch decoder Transformer on (a + b) mod 113, trained far past zero training loss. It memorizes at step 81 and doesn't generalize until 1,350 — a 16.7× gap. Turn weight decay off and generalization never arrives at all: training stays perfect for the remaining 39,920 steps while held-out loss ends at 19.9, against 4.73 for a model that knows nothing.
Own the full GPT stack: a from-scratch GPT (≤~10M params) plus a data-vs-model ablation on a free T4. Study note written; reproduction underway.
In regulated aircraft maintenance, a wrongly-signed repair isn't a bug: it's a failed audit or a grounded plane. The work is making the platform enforce who's allowed to act, and keep the proof.
Lead backend engineer. I own the .NET modular monolith behind ~10 React/TypeScript apps, and I'm the one shipping the LLM and computer-vision features layered on top.
Business-critical flows fail quietly when an event slips between services. This role was about building Go foundations solid enough that none did.
I built the Go services and event pipelines the rest of the product depended on, and made them the boring, dependable kind.
Support and monitoring teams were working without a shared record of what was actually happening. The job was to build the internal tools that gave them one.
My first production work: a helpdesk system, a monitoring project I ran, and a point-of-sale backend. All of it shipped while learning how the people using them actually worked.
Before any of it carried a title, the question was whether I could take a real product from nothing to shipped. This is where I answered it.
React / Express / Node.js builds: a facial-analysis web app, an online training platform, and a news portal.
