Pikki: a live, paid Chrome extension

Context

Job hunting runs on bad leads. A large share of postings are ghost jobs or stale reposts, employment scams are rising, and most applicants have no route to a referral. Pikki is a Chrome extension that treats those as product problems, not content problems.

It is live at pikki.synthworx.com, with a free tier and three paid plans ($19, $39, $79 per month). It runs in the user’s own signed-in Chrome.

The problem

The product had to do four things at once, on consumer hardware, without becoming another source of fabricated applications:

1. Tell a real job from a ghost job. A posting can be an old repost, a resume-collection ad, or an outright scam. The user needs evidence, not a score.
2. Find a way in. Referred candidates are hired at a much higher rate than cold applicants, but most seekers do not know who to ask.
3. Apply honestly. A generated resume that invents a title or a date is worse than no application. The system must only reword what is already true.
4. Keep the data local. Resumes, answers, jobs, and applications live on the user’s computer. The server passes model requests through and keeps counts, not content.

Architecture

Pikki is split between the extension, a local data layer, an on-device model path, and a server that brokers model calls. The extension owns the user’s session and the page; the server never holds the user’s resume.

Simplified. CAPTCHAs and account creation always wait for the user.

The key decisions:

• Local-first. The database is on the user’s machine. Model requests are proxied; the text is not stored. The optional profile backup is unticked by default in the UK, EU, and EEA.
• On-device where it fits. Searching, scoring, the fake-job check, people search, Gmail status, and outreach drafts run on Chrome’s built-in Gemini Nano where the machine supports it. A user can add their own DeepSeek key to unlock the rest at no model charge from Pikki.
• Evidence, not verdicts. A ghost-job or scam flag quotes the page or record that produced it. The user can check the reasoning.
• Approval gates. Pikki never solves a CAPTCHA, never creates an account without approval, and never submits twice. The user decides whether Pikki clicks the final Submit.
• No invented facts. Tailored resumes reword existing content. Anything the system cannot answer from the profile comes back as a question.
• A self-updating native helper. Because the extension is installed outside the Chrome Web Store, a small native helper updates it from a signed, checksummed download. The helper takes no address, path, or command from anything except Pikki.

What I did

• Designed and built the product end to end: extension, local data model, model routing, job-board integrations, billing, and update path.
• Built the ghost-job and scam checks with page-level evidence.
• Built the referral search and outreach drafting flow.
• Built the application preparation and submission flow with approval gates and a no-fabrication rule.
• Built the local-first privacy model and the public privacy page.
• Shipped it as a live paid product with three plans, not a demo.

Results

This is a product case study, not a client engagement. There is no client metric to report. The proof is that the product is live, priced, and used — with a privacy page that describes exactly what leaves the machine.

What it proves

A working product with pricing and a privacy page is stronger evidence than a deck. It also forces the unglamorous parts — install, update, billing, data residency, approval gates — to be real. Those are the parts that decide whether an AI product survives contact with users.

The product is live at pikki.synthworx.com. The source stays private; the architecture and demo are in pikki-showcase.