Software Engineer, Verification Fleet

What You Will Own

• The robot fleet, end to end. The browser automation that drives real checkouts, the classifier that sorts ~500K merchants into 50 to 80 platform families, and the scheduler that decides which stores to test and when.
• Agents write much of the code; you own the design, the failure modes, and the verdict on what ships.
• Coverage as your number. Machine-tested checkout today reaches 21% of non-Shopify merchants; you own the line from there past 80% across roughly 500,000 stores.
• Fleet economics. Cost per verified checkout, held below the commission each check protects. You make the spend legible and make the fleet earn its keep, store by store, rather than making it small.
• Anti-bot navigation. The evolving contest with fingerprinting, rate limits, and challenge walls — navigated at scale without breaking the store or the law.
• Instrumentation that proves it. Dashboards and ledgers that show, for any claim, when it was last tested, whether the robot really reached the cart, and what the check cost. Correctness you can watch, not correctness you assert.
• The number, co-signed. Within your first quarter you co-sign a seat charter — the model we run for senior operators. It names one machine-checkable number that proves the seat works (machine-tested checkout coverage is the obvious one) and writes down what you decide freely versus what you propose for the founder to sign. You own a number, not a backlog.

The System You'll Need to Model

• The checkout maze. Every e-commerce platform breaks differently. The code field hides behind a different click on Shopify, Magento, WooCommerce, BigCommerce, and a thousand custom carts. A classifier has to sort roughly half a million merchants into 50 to 80 platform families, so one robot recipe covers thousands of stores instead of one.
• Fleet economics. A robot that costs more to run than the commission it protects is a loss. Every check spends compute, proxies, and time. You are always trading how often you re-test against what the test is worth — the fleet has to earn its own keep, store by store.
• Coverage-accuracy tradeoff. Pushing coverage from 21% toward 80% means testing messier, stranger stores where a robot is likelier to misread the cart. More reach and more certainty pull against each other, and holding both as you scale is the whole game.
• Evasion versus detection. Stores and their anti-bot vendors do not want to be automated. Fingerprinting, rate limits, and challenge walls move constantly. You navigate them without breaking the store or the law, and the ground shifts under you every month — this is a moving contest, not a fixed integration.
• Verdicts need tests; tests need verdicts. You build beside the seat that owns the scoring science — the machine verdicts that decide what we claim is true. Those verdicts are only as good as the checkout evidence your fleet produces, and your fleet only knows what to test because their scoring shows where the truth is thin.

Who You Are

• Reason in invariants, failure modes, and tradeoffs. Handed a checkout flow you have never seen, you can sketch the three ways it will break before you write a line.
• See the platform family behind a one-off store, and the shared recipe behind a hundred one-off stores. When a robot fails at 2 a.m., your first question is structural: what class of store did we just discover?
• Move fluidly between architecture and shipped code — a classifier design in the morning can be a deployed test by night — and you are as comfortable deciding what to build as how.
• Treat agents as leverage you verify, not autocomplete you trust: you can point at a system you shipped, name the hardest failure you personally diagnosed in it, and say what you changed.
• Have built browser automation, web scraping, or crawling systems at real scale, and you operated them in production — you know what a fleet of headless browsers does to your infrastructure bill and your on-call sleep.
• Reverse-engineered a site that did not want to be automated, and won.
• Playwright, headless Chrome, proxy rotation, and queue-backed job systems are familiar ground; Node.js and Python are daily tools.