Learn

The origin story: forking a viral-but-abandoned Interview Coder project into Ezzi, an invisible desktop overlay that feeds you AI-generated solutions during coding interviews without the interviewer noticing. Why it is now open-source and where it is headed.

Inheriting a viral, AI-generated Electron codebase meant security holes, `any` types everywhere, 700-line files, and client-side credit logic. Real code examples and the patterns to watch when maintaining someone else's vibe-coded project.

LeetCode-style interviews do not predict job performance, yet the industry keeps doubling down. A comparison of common formats, why each falls short, and why tools like Ezzi are a symptom, not the disease.

A technical deep-dive into making an Electron overlay truly invisible to screen capture: how each OS captures windows, the BrowserWindow properties that matter, and the platform-specific edge cases (including macOS Zoom).

I built an interview-assist tool, so I am not neutral about the ethics. An honest walk through the strongest arguments on both sides of using AI during interviews — and where I have actually landed.

Run Ezzi entirely on your own stack: clone the repo, stand up a single Fastify endpoint, wire it to Claude, and own every request and API key. Full private setup in under an hour.

Companies run AI at every hiring stage — job descriptions, resume screening, question banks, automated grading — while the cheating conversation lands entirely on the candidate. Flipping the camera on a field that was never level.

Against the 'charge more' religion, with the actual P&L of a tiny indie product: 33 paying users, ~$70 a month, and a cheapest plan of $2.99 for a day of access. The margin math, and why the default indie-hacker playbook is aimed at a buyer that isn't mine.

You don't beat the interview by seeing 500 problems — you beat it by recognizing the 15 patterns wearing 500 costumes. Why grinding burns your evenings without building pattern recognition, and a repeatable approach for anyone with an interview next week.

A technical deep-dive into Ezzi's AI vision layer: a cheap Sonnet model reads the interview problem off a screenshot while an expensive Opus model solves from the transcript and never sees the image. The real production failure modes — split honestly between the ones that got fixed and the ones that are still just limitations.

A guardrail built to filter out system-design and behavioral questions started refusing the simple coding tasks it was supposed to solve. A plain 'flatten this JSON' task fell into the OTHER catch-all and got rejected — the taxonomy was missing a bucket, and the fix was a first-class CODING type, not a smarter prompt.

Ezzi ships as a free, open-source app you can self-host and as a managed Cloud backend you rent by the day. A straight comparison of what each really costs in money and in time, where self-hosting genuinely wins — and why the API bill is almost never the thing that should decide it.