ResumeMatch Coach

Scores my resume 74/100 ("Good fit") and splits keywords into matched vs. 8 missing.
Real strengths from my experience, plus concrete, honest fixes for each gap.
More rewrites adapting my real experience to the job's requirements.
Anticipates tough questions, like whether 99.32% accuracy could be data leakage.
My real bullets rewritten in the job's language, before → after, with keywords added.
Likely interview questions, why they're asked, and how to answer from my resume.

Inspiration

As a new grad student applying to US summer internships, I kept hitting the same wall: every job description is different, applicant tracking systems (ATS) filter on keywords, and I never knew whether my resume actually fit a role or how to fix it. Career centers are booked for weeks. I wanted a recruiter in my pocket that gives honest, specific feedback in under a minute.

What it does

Upload your resume as a PDF and paste any job description. ResumeMatch Coach then:

Scores your fit (0–100) with a strict, recruiter-style verdict and summary
Finds ATS keywords: the ones your resume already shows, and the important ones it's missing
Lists strengths and gaps, with concrete, honest ways to address each gap
Rewrites your resume bullets for that specific job (before → after) without ever inventing experience, tools or numbers
Prepares you for interviews: 5 likely questions, why they're asked, and how you should answer using items from your own resume
"Ask the Recruiter" chat for follow-ups like "How do I explain this gap?" or "Write my cover letter opening," with your resume and the job description in context
Downloadable report of the full analysis ## How we built it
Gemini 3.8 Flash through the Google GenAI Python SDK
Native PDF understanding: the resume PDF goes straight to Gemini, with no text-extraction step, so layout and formatting are preserved
Structured output: a Pydantic schema (fit score, keywords, gaps, rewritten bullets, interview questions) enforced with Gemini's JSON response schema, so every result renders reliably in the UI
Prompt design: a senior-recruiter system prompt with strict honesty rules (no inflated scores, no fabricated experience)
Resilience: automatic retry with backoff, plus fallback across Gemini models when one is overloaded or retired
Gradio web app (report, bullets, interview prep and chat tabs) running on Google Colab, shared through a public link

Challenges we ran into

Model availability: our first model choice had been retired for new API users (404), and under hackathon load the newest model often returned "high demand" errors (503). We built a fallback chain that retries, then switches models automatically, so the app keeps working.
No local setup: we were on a loaner laptop with no Python, so we built and ran everything in Google Colab.
Honesty vs. helpfulness: LLMs love to "improve" a resume by sneaking in skills the candidate never had. We wrote strict no-fabrication rules into the prompt and added a "keywords added" field to every rewrite so users can verify each change. It's not perfect yet: in testing, one rewrite added "SQL" to a bullet that didn't mention it, so the app is a coach, and the user stays the final editor. ## Accomplishments that we're proud of
A working end-to-end app built within the hackathon
Feedback that is specific to the person: it cites real resume items instead of generic advice
Reliable, structured results thanks to schema-enforced JSON output
Tested on my own resume against a Data Science internship job description: scored 74/100 ("Good fit"), found 8 missing keywords (pandas, scikit-learn, PyTorch, dashboards, GCP, AWS, A/B testing, LLM), rewrote 5 bullets in the job's language, and generated 5 interview questions tied to my real projects, including a sharp one asking whether my 99.32% accuracy could be data leakage ## What we learned
How to combine Gemini's multimodal PDF input with structured output
Prompt engineering for honesty constraints
Building for real-world API limits: rate limits, overloads and model deprecations