Software engineer intern at a fintech company; mechanical engineering senior at Kennesaw State. I direct AI coding agents to ship production software and audit everything they produce. Atlanta area; happy remote or on-site.
Numbers I can prove, from production systems:
Any employee clicks Feedback; an agent reproduces the bug, fixes it, and opens a pull request. A human decides what ships. First six reports: five merged PRs, four the same day.
Bug reports are untrusted text, so the pipeline is hardened against prompt injection and fails closed.
A skill that turns review into a gauntlet — 41 agents, three adversarial passes, two fixers. Nothing ships on one model's opinion.
A barbecue restaurant runs payroll on this: PIN-entry terminal, overtime rules, and Excel exports that match the payroll processor's template byte for byte. 226 tests, and one invariant I'm fond of: the export file exists if and only if an approved pay period does.
A weekly paycheck allocator built for one real user. The domain model strictly separates pre-tax and post-tax dollars, so tax math and budget math can never contaminate each other.
Extended an open-source MCP server so Claude can drive my CAD — modeling, drawings, batch export. I'm an ME student; my homework tooling writes VBA macros for me.
I'm a mechanical engineer by training: SolidWorks and AutoCAD, 3D printing, reverse engineering, and coursework through machine dynamics, thermodynamics, heat transfer, and strength of materials. That training shows up in my software.
The habits transfer. Unit checks, tolerance thinking, documented assumptions. A free-body diagram and a reconciliation harness are the same instinct: don't trust a result you haven't checked against the physics.
I've been your user. For teams building software that touches the physical world — robotics, manufacturing, CAD and simulation tooling — I'm the engineer on the other side of your product, and I can translate.
The proof is above. The SolidWorks project in my work list is the two sides meeting: AI agents driving real CAD through an MCP server I extended, generating VBA a machinist could read.
Requirements are a spec. Engineering taught me to turn a vague ask into testable design inputs before building. That's the same job as writing a good ticket.
Start with a premortem. Assume the plan already failed; find out why before committing.
Verify at the seam. Two correct components can still be wrong where they meet. Call the deployed thing on real data.
A green test suite is not evidence. An audit once found 11 real bugs behind 2,673 passing tests. Break the code, watch the test go red, restore it.
Reproduce before shipping a hypothesis. Four wrong diagnostic PRs taught me this one. My bug-fix agent carries it as a hard gate.
Make wrong answers loud. A system that can't be correct yet should warn, never guess.
Write down what cost you a day. The same day is never lost twice.
Tell me what you're building.
Email is the fastest path. Say what "done" looks like and I'll respond with something useful.