Inspiration

After university, many of us face our first real interviews without knowing whether our CV fits the job or how an interview will feel. Practice tools usually ask generic questions from a fixed list. We wanted something that behaves like a real HR interviewer: it reads your CV, knows which job you're applying for, and is honest with you at the end.

What it does

IntervYou is a web app for practicing a job interview in English or Arabic.

Choose the language, upload a PDF CV, and type the job title. If the file isn't a CV, the app says so and doesn't start.
An AI interviewer asks six questions, one at a time: two based on your CV, two role-specific scenarios, one about what you would contribute to the company, and one about where you see yourself in two to three years.
If an answer is vague, it asks a follow-up (at most two per question). It gives no feedback during the interview.
At the end you get a report: an overall score, six scores out of 10 with reasons, strengths, areas to improve, how well your CV fits the job, and three suggestions to improve your CV. In Arabic mode the interface and report are right-to-left.

The score is a practice estimate, not a prediction of a real interview result.

How we built it

Plan first. We used the Devpost Learn Skill Pack to write scope.md, prd.md, and spec.md before any code, and added short change notes when the design changed. They are in the devpost/ folder.
Stack: Python, Streamlit, pypdf for local CV text extraction, and the Google Gemini API (gemini-3.5-flash-lite).
Rules in code, not just prompts. Python enforces the six questions in a fixed order, the two-follow-up limit, answer and job-title validation, and a local "is this a CV?" check. Gemini writes the questions, but the code decides what is allowed.
Arabic support. We normalize extracted PDF text (Unicode NFKC), because Arabic PDF text can come out as presentation-form letters that break matching. The interface switches to right-to-left, and English terms inside Arabic sentences are isolated so they display correctly.
Privacy. The app doesn't save CVs, chats, or reports. The API key stays in a Git-ignored settings file. Only the CV text and conversation go to Gemini, so we test with fake CVs only.
We deployed a test version on Streamlit Community Cloud for teammate testing; it may pause if the free API quota runs out.
76 offline tests, plus manual tests with fake CVs of different levels and fields in both languages.

Challenges we ran into

Scope. We cut accounts, saved history, voice, and PDF export. Arabic started on our "later" list; we added it at the end on a separate branch so the English version stayed safe.
Arabic PDFs. Extraction returned presentation-form letters until we normalized the text. Scanned CVs have no readable text and we did not add OCR, so the app shows a clear error.
Wrong files. A lecture file was accepted as a CV during testing, so we added the CV check before any AI call.
A realistic interviewer. We tuned the instructions so it follows up on vague answers without repeating itself or leaking feedback, and changed the questions to match real interviews.
Free usage limits. AI tools and API quotas ran out while we worked, so we planned in short steps and added automatic retries for overloaded responses.

Accomplishments that we're proud of

One complete flow in two languages: upload a CV, take a six-question interview, get a report, restart.
An interviewer that stays in character and gives no feedback until the end.
An honest report that says plainly when a CV doesn't fit the job.
Interview rules enforced in code and covered by tests.

What we learned

Planning first made the AI agent far more useful, because we could check the app against written requirements.
Constraints belong in code: prompts guide the model, but rules like question order must be enforced by the program.
Testing with messy real input (a wrong file, an Arabic PDF) found problems our planned tests missed.
Be upfront about data: the app tells users that their CV text and answers are sent to Google Gemini.