I sold and onboarded every SwipeBy customer myself. Now I build AI systems and the tests that check them.
Looking for solutions engineer, customer engineer, deployment strategist, go-to-market engineer, and applied AI engineer roles in Chicago or remote.
How I check AI agents
Agents should not be able to grade their own homework.
If an AI agent can chase its own score, the score stops meaning anything. That is Goodhart's law: when a measure becomes a target, it stops being a good measure. Below is how I keep agents apart from the code that scores them, and how I test the product every night. It is part of Shayke, an applied learning project I built with AI coding agents in my own time. It has no customers.
Keeping agents apart from their grader
The agents do the work. A separate program, the scorer, grades it.
Click a move to see what happens when an agent tries it.
Import the scorer: stopped by tests.
Write a result file: allowed, the only path across.
Edit a golden test case: by design, they live in the scorer's own repo.
Open a score file: known gap, nothing at runtime stops it yet.
What this stops
An agent shaping its work around the scoring code. Tests fail if an agent's code imports the scorer or names where scores are stored.
A test run every night, the evaluator
The evaluator also scores agent outputs using the scorer's checks. Back to the wall
19 of 19 scenarios passedLatest run, on hard checks, 20 Oct 2026.
What I checked, and what I have not finished
Checked 20 Oct 2026
Nightly evaluator, latest run: 19 of 19 scenarios passed on its hard checks.
Checked 12 Oct 2026
Product test suite: 10,446 passed, 32 skipped.
Checked 20 Oct 2026
Quote check: 39 tests passed. It is public and runnable in shayke-public, folder lib/quotable_span/.
Built, not merged
A judge that must quote exact evidence for its verdict. It lives on a branch.
Known gap
Nothing at runtime stops an agent opening a score file. Read protection is convention plus tests. It is the next thing I would close.
Status
The agent fleet is paused. The nightly evaluator and a few scheduled jobs still run on one machine.
Selling, then building
SwipeBy, Sales Manager and Account Executive
March 2025 to February 2026
12 multi-unit deals and about $60K in annual recurring revenue (ARR) in my first three months, a company record, and a promotion at three months. I moved to Chicago to open the market here.
Nobody asked me to, but I built a lead pipeline in Python with the Google Places API and a language model that loaded restaurants and wrote up their pain points, replacing manual CRM entry. AI wrote much of the code; I designed it, directed it and ran it. Code on GitHub.
I also built a shared AI assistant that put research on each location in reps' hands before a visit, and grounded it so every claim had to come from its knowledge base. My reps used it without being told to.
Background
Seven years as a Senior Constable with Victoria Police: deciding under pressure with incomplete information, then standing behind the decision.
A year as Australian Army tank crew.
Division I football at the University of Illinois, and a BS in Management from Gies.