Pick your own song above, or .
I build tools that catch flaky browser tests and AI answers that make things up.
I work in TypeScript and Python on the bugs that are annoying to pin down, like a test that fails 1 run in 5 or a RAG answer that isn't actually in the PDF. I also send fixes upstream: seven are merged: six into Google's Puppeteer and Chrome DevTools MCP, and one into Hugging Face's sentence-transformers.
Six projects, each with something you can open and try. Zeitaku and Junban are the two I'm most proud of.
A CLI that finds waitForTimeout sleeps in Playwright tests and replaces them with a wait that checks when the network is actually done (using the Chrome DevTools Protocol). I parse the test files with the TypeScript compiler so it doesn't touch comments or strings.
Finds the exact order of API responses that makes a flaky Playwright test fail. It holds each response in the browser and lets them through in different orders, then shrinks the failure down to one rule like "fails when /api/cart arrives before /api/user" and writes a test that fails every time.
Checks if a RAG answer is backed by the retrieved text. My first version used cosine similarity, but it scored wrong answers high if they used the same words, so I switched to an NLI cross-encoder (DeBERTa-v3).
Tests a RAG app without knowing the right answers. It changes the input in ways where we know what should happen: reword the question, add an unrelated document, delete the document with the answer, change a number in it. Then it checks the answer reacts correctly and shrinks each failure to the documents that cause it.
Upload a PDF and ask questions about it, scanned PDFs too. It only runs OCR on pages that have no text, which made normal PDFs way faster. If nothing in the doc matches, it just says it doesn't know instead of asking the LLM.
My own idea: it watches which app and site you're on and whether you're typing or just scrolling, asks you every 20 minutes how focused you really are, and learns your personal "drifting" pattern from those answers. It nudges you when you drift and draws a daily focus report. Everything stays on your computer.
Fixes I sent to tools I use: six merged into Google projects, one into Hugging Face, and a bug I found in Microsoft's Playwright that their team fixed.
I assumed every waitForTimeout should go. Then I read through real test suites (Documenso, Supabase, n8n and a few more) and about 1 in 6 were there on purpose. They wait before checking that something didn't happen, so removing them would make the test pass when it shouldn't.
So now Zeitaku looks at the next few lines after each sleep before changing it. Out of 240 sleeps it found, 196 were safe to replace and 44 it leaves alone.
// documenso: we must NOT navigate away await page.waitForTimeout(1000); await expect(page).not .toHaveURL(/\/admin\/users\/\d+$/); $ zeitaku scan ./tests Total waitForTimeout: 240 replaceable: 196 intentional: 44 (left alone)
What I learned building these tools, with the real numbers.
Thank you for coming here.
I'm looking for my first engineering job, ideally at a startup working on testing, browser tools, dev tools or AI. Remote or relocation both work for me. Email me or DM me on X, I usually reply the same day.