Rizzume: tailor your resume to any job with Gemma, without your data leaving your laptop

Subhasish Das — Sun, 18 Oct 2026 17:08:21 +0000

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle

What I Built

Every job application wants a slightly different resume, and tailoring one by hand is slow. The quick fix is pasting your whole history into a chatbot, which hands your phone number and work history to a server you don't control.

Rizzume keeps your projects and skills in one profile. You paste a job post, and a Gemma model running on your own laptop picks the most relevant projects, rewrites the bullets in the job's wording, and gives you a one-page, ATS-friendly PDF.

Who would pay: students and early-career developers who apply to many roles, and college placement cells that can't send student data to third-party AI tools. I would sell it as a small one-time student pack, somewhere around $3 to $5. A local model costs nothing per resume, so a small flat price is still profit. A paid API charges per request, which makes a cheap one-off product hard to run.

Demo

Note: the video shows the real interface. For the recording I used a scripted stand-in reply for the model step, so the resume text in the video is not Gemma's output. The real app runs on Gemma through Ollama, as described below.

Code

Nikshan63 / Rizzume

Rizzume (formerly ResumeFit)

Store your projects and skills once. Paste a job description. Get a one-page, ATS-friendly resume tailored to that role The model is an open-weight Gemma model running locally through Ollama, so your personal data never leaves the computer running the app.

How it works

You enter your details once (name, contact, skills, projects, education, optional photo). They are saved as a profile.
You paste a job description.
The local model picks the most relevant projects, rewrites the bullets in the job's vocabulary, and orders your skills.
Guardrails check the model's output before you see it: invented skills and projects are discarded, and any bullet or summary containing a number that is not in your stored facts is dropped.
You edit anything you like, check the ATS panel, and download a one-page PDF.

The page also shows a rizz meter (how many of the job's tech skills…

View on GitHub

How I Built It

Open-source AI: Gemma (gemma3:4b), an open-weight model, served locally by Ollama. FastAPI backend, ReportLab for the PDF, a single-page vanilla JS interface.
The model only sees my stored facts and the job text, and must answer in JSON: chosen projects, ordered skills, rewritten bullets, a short summary.
Guardrails run after the model. Unknown projects and skills are discarded, and any bullet or summary containing a number I never wrote is dropped. If the model returns garbage, a keyword-matching fallback picks the projects.
Rizz score: how many of the job's tech skills you have, computed by code, not guessed by the AI. A gap list shows the skills you're missing instead of hiding them.
ATS-friendly on purpose, and checked: one page, single column, real text, standard headings. After each run the app reads its own PDF back like a parser would and reports whether the name, email, phone and headings come through.
Optional passport photo, cropped locally. It is a separate image in the corner, so the text layer is unaffected. Required fields (name, email, phone, location, a skill, a project, education) must be filled before a resume is built.
Live counter: shows how many people are using the running copy and for how many different job posts. It stores only counts and hashes, never resumes or job text.
46 automated tests cover the guardrails, photo handling, ATS check, validation and the counter.
A real bug: my first version crashed with an Internal Server Error on some job posts, most likely because the model returned empty or oddly shaped fields. I fixed it by cleaning up the model's output before using it, and by showing the real error on screen instead of a vague message.

I ran everything on a MacBook Air with no GPU server. I did not run a formal benchmark, so I am not claiming a speed number.

Why Does Open Innovation Matter?

Privacy is the product. A resume is personal data. With an open model running locally, nothing leaves the machine.
No cost per resume. That is what makes a cheap, flat-price product possible.
Control. Because I run the model, I could force JSON output and add checks around it. The guardrails that remove invented skills and numbers only work because I control the whole pipeline.
Where it worked well: a 4B model on a normal laptop produced usable structured output end to end. Picking projects and rewriting bullets from my own facts, with the guardrails behind it, was enough to give a clean one-page resume.
Where it fell short: a small model is less predictable. It broke my first version with badly shaped output, which is why I added the cleanup step and the guardrails. It also runs slower than a hosted API would on a laptop, so this is a tool for people who value privacy over speed.

My First User

No one besides me has used it yet. I built the live counter so the first real users can see themselves counted.

Prize Categories

Gemma (the app runs on gemma3:4b through Ollama). I did not use MongoDB Atlas or a .tech domain.

What I would build next

Match job requirements to projects with embeddings instead of keyword matching.
Parse a job post from a link or screenshot.

DEV Community: Subhasish Das