JobbyAIresume scoring · job match · interview prep

The original question was “what would a minimum useful career-coach look like if it had access to a capable LLM and nothing else?” The answer we arrived at was three tightly-scoped tools that a job seeker would reach for in the same session: score my resume, rank me against these postings, prepare me for this interview.

Deliberately not a chatbot. Chatbots are easy to build, easy to show off, and a poor fit for someone who is already anxious about the conversation they're preparing for.

• Time-to-first-suggestion under 2 seconds. Anxious users do not wait.
• The product has to feel trustworthy: every score needs a reason, not a vibes-based number.
• Keep the dependency surface small — this is a side project that has to be maintainable in an afternoon a month.
• Free to try, transparent about costs, no dark patterns on upgrade prompts.

A Next.js app on Vercel, a single well-prompted Gemini call per flow with structured output validated against a schema, and a Tailwind-driven UI that is deliberately quiet — white space, legible type, one accent color per section.

Each flow is a pure function of pasted text in, structured JSON out, React component render. No session state, no “chat history,” no stored user data unless the user signs in. The code surface is small enough that you can read the whole thing in a sitting.

• Prompt stability. Gemini's structured output mode was still young when we built; one in a hundred responses would return near-valid JSON that had to be coaxed, not rejected.
• Job-match ranking is almost entirely a prompt engineering problem, not a classical ML problem. Getting the model to produce a defensible rationale without hallucinating matching requirements took more iteration than everything else combined.

Running on Vercel, open source under MIT. Users report the resume scoring and interview-prep flows are the ones they come back to — the interview prep was almost cut in week two and turned into the most-used feature by week six.

I'd ship a structured feedback loop from day one — a thumbs up/down with an optional “why?” — and feed it into prompt iteration. The version we have works because we read every piece of user feedback by hand; that doesn't scale past a weekend project.