Chrome extension
slayply
A Chrome side panel that turns a job description and its application questions into answers written from your real background, with your choice of model.
What it is
Slayply lives in Chrome's side panel. On a job page, you capture the description and the questions. It checks how well you fit, picks the right résumé, and drafts answers, a cover letter or a message. You bring your own key for Claude, OpenAI or Gemini.
It never fills in or submits anything. You read, edit, copy and submit yourself, every time.
Why I built it
Job applications ask the same questions in slightly different words, and generic AI drafts read like generic AI drafts. Worse, they invent things: a metric you never hit, a tool you never used. I wanted drafts that can't.
- 11 daysfrom first commit to v1.1.0, 29 Oct to 9 Nov 2026
- 150+commits, through pull requests and CI
- 1,100+test cases, mostly against stubbed providers
- 23written architecture decisions
The idea that matters: claim checking
You keep one profile of real facts, and each fact has an ID. Every sentence the model writes about you has to cite one. A deterministic checker then highlights any sentence with no citation, so an invented claim is flagged before it reaches an application.
How it's built
- Side panelWXT, React, IndexedDB
- CoreZod schemas, prompts, claim checker
- RelayHono, stateless
- Claude, OpenAI, Geminione adapter each
It's a monorepo with six packages: the extension, the relay, a static site, a shared core, the provider adapters and a UI kit. Your profile, résumés and keys stay in the browser. The relay keeps nothing, and there is no telemetry.
How I built it
I wrote the spec and the architecture decisions first, then worked with coding agents inside a harness: a handoff pointer file the agent reads at the start of every session, tests written before the change, and hooks that block risky actions. I describe the whole setup in How I use AI in my production workflow.
Where it stands
It has had one real-world test: answer generation worked end to end with a real Gemini key, slowly. A fit check on a slow thinking model ran past the relay's time limit, and I've added progress and retry for it. The Claude and OpenAI paths were written from the vendors' documentation and haven't been run live. If you have credits and ten spare minutes, an issue saying what broke, or that it worked, helps a lot.
What I'd change
I'd test against real providers earlier. Almost everything is verified with stubs, which proves the plumbing and says nothing about how real models behave on real job pages.
Need someone who owns a feature from the first call to the last deploy?
I'm open to senior software, full-stack and forward deployed roles, in India or abroad. A 15-minute call is the quickest way to see if we fit.