Turning Job Discovery Into a Review-Ready Application, One User at a Time

A common answer is the auto-apply bot: a tool that sits on a shared candidate pool and fires applications at listings on the user's behalf. That trades control for volume and often sends generic, low-quality applications the person never reviewed. This platform takes the opposite stance. There is no candidate pool an employer draws from, and the platform does not choose jobs or send anything automatically. Every action runs inside one authenticated user's account, and the person stays in charge of which jobs to pursue and what actually gets sent. The problem to solve was narrow and real: remove the repetitive preparation work without removing the applicant's judgement.

The constraints were deliberate. Human-in-the-loop is non-negotiable: the platform prepares a Gmail draft and never sends it automatically, so the applicant always reviews and clicks Send. The AI may only use information the user supplied or approved; it must never invent qualifications, employment history, skills, certifications, achievements or metrics, and must not add skills just to inflate keyword overlap. Each user's data is strictly separated: a user only ever touches their own profile, searches, selected jobs, generated documents, Gmail drafts, credits and tracked applications, never anyone else's. And because generation costs money to run, the cost had to be transparent: the user should see the credit price of an action before the AI starts.

Preparation is where the AI does the repetitive work, and where credits apply. The user manually selects the jobs worth pursuing and clicks Apply Now on each. For every selected job, the platform generates a separate, tailored package: a résumé draft aligned to that job description, a cover letter for that role and employer, and a customized application email, so two applications from the same user can differ because each is built for a different opportunity. It then uses the user's Gmail integration to assemble a draft, a job-specific subject, the application email as the body, and the résumé and cover letter attached, and saves it as a draft. It does not send. The user opens the draft in Gmail, reviews the employer, role, recipient, subject, body and attachments, edits anything, and manually clicks Send. Once sent or marked applied, it lands in that user's application tracker.

Generation is gated by credits so the economics match the value: discovery stays free, and the computationally expensive preparation is what users pay for. Each user gets a daily free-credit allowance to try or continue the workflow; when it runs out, they can purchase more. The interface shows the credit cost before generation begins, so the user understands the price before invoking the AI.

The Apply Now workflow first checks credits. If the balance is short, the user is routed to purchase; if it is sufficient, the required credits are deducted and generation begins. The AI generation step takes the user's profile and the selected job description and produces three tailored drafts, a résumé, a cover letter and an application email, which flow into the user's Gmail as a single draft with the documents attached. The user reviews and edits the draft, manually clicks Send, and the application is recorded in their tracker. Two design lines run through the whole system: human-in-the-loop, so the AI drafts but the user always sends, and per-user isolation, so profiles, searches, generated documents, Gmail drafts, credits and tracked applications belong to one account and are never shared into a common candidate pool. The stack stays deliberately boring so the moving parts remain legible: a React front end, a Node.js backend, Supabase for data, and Vercel for hosting. The Source Pool is built to grow: each new third-party job-portal integration plugs into the same aggregation and scoring path, so adding a portal widens coverage without changing the workflow.

What I would strengthen next builds on that base: make credit transparency richer and unmissable (cost per action, refund on failed generation, whether edits or regeneration cost more, when the daily 50 reset), so no generation ever surprises a user on price; add match explanations that clearly separate confirmed strengths from missing or unconfirmed requirements, so the score is actionable rather than a number; make the candidate-provided-facts guardrail visible with clear provenance on every generated line; widen the Source Pool with more portal integrations; and deepen the application tracker into a genuine pipeline. The lasting idea is a job-search assistant that does the repetitive preparation at scale while the applicant keeps every decision that carries their name.