About
How InterviewGym works.
You upload your CV, the job description, your interviewers' LinkedIn profiles, and the company's two latest annual reports. We try to parse documents in your browser first — when that works, the original file never leaves your device. When your browser can't (older iPhones, scanned PDFs, large files), the file is briefly processed on a stateless server route that does not store it. Either way, only the extracted text is kept in your session, in your browser. The text is then forwarded to the K2 Think V2 API to generate your panel, your question bank, and your post-interview report.
We do not store your documents or your answers. There is no database, no logs of request bodies, no file storage. Each backend route processes the request and returns the result; nothing about your session is persisted server-side. Exporting the report as a PDF uses the same pattern — your report text is sent to a stateless rendering route that returns the PDF and discards the request.
We do collect anonymous, aggregate page-view counts via Vercel Analytics — no cookies, no cross-site tracking, no PII, no link to your documents or your interview content. It tells us roughly how many people are using the product. That's the only thing we measure.
Your session — the brief, the questions, the answers, and the final report — lives only in your browser via IndexedDB. Closing the tab and clearing site data wipes everything.
The K2 Think V2 API is operated by a third party (MBZUAI / their hosting provider). What that provider retains in their API logs is governed by their terms; we do not transmit identifiers tying a request to you beyond the document text itself.
How the brief is built
The brief is not a single round-trip with an LLM. It is a multi-stage pipeline.
The document parser pulls structured anchors out of your CV — the employer names, the role titles, the certifications, the geographies, the metrics — and passes them to the model as the authoritative list of names it is allowed to use when referring to your career. The model is told plainly: invent a placeholder, fail the entire output. A server-side validator then runs every generated question against those anchors and drops anything that isn't grounded in your actual history.
For Full Session, the company name drives a recent-news search before the brief is generated, so the brief reflects what's been on the company's plate this quarter — moves, deals, segment shifts. Each interviewer is matched to the questions a person with their actual background would naturally ask.
The technical questions are bound to specific claims in your CV. A CFA charterholder gets a portfolio-construction question with named tensions, not "tell me about your CFA." An ISO 45001 lead auditor gets pushed on where the management-system theory broke down in practice. A practitioner who claims a deal volume gets probed on a specific deal's structure.
How the grading works
Every answer is graded against multiple independent quality dimensions, with anchored 1-to-5 scoring on each. Scores stay independent — nothing is averaged into a single number, because averaging a panel's judgment is exactly the move that produces the bland feedback most interview-prep tools generate.
The grader is told to be tough, not polite. A score of five is rare. Recognisable filler — placeholder text, repetition, gibberish — locks every dimension to one and the report says so plainly. Polished textbook answers that miss the substantive practitioner signals score lower than rough answers that hit them.
When a specific dimension lands weakly, the next interviewer turn is an adaptive follow-up that pushes on that exact thread, in the voice of the same panelist. Topics don't shift while a thin answer is left unchallenged.
The closing question — what you ask the panel back — is graded as first-class signal alongside the answers, with its own dimensions for company-specificity, forward-orientation, and absence of red flags.
Why K2 Think V2
K2 Think V2 is a reasoning-native model. Multi-criteria evaluation stays as structured reasoning per dimension rather than being collapsed into a single conversational take.
Chat-style models tend to flatten this — given an answer, they produce one paragraph that averages the dimensions into a polite summary. K2 holds the dimensions distinct, scores each on its own evidence, and the report synthesises from the structured grades rather than rewriting them. The choice of model is deliberate.
What InterviewGym is
A serious, single-session interview-training app. Two modes: Sparring for quick CV-only practice, and Full Session for a panel interview built from your real documents and recent news on the company.
What InterviewGym is not
It is not a chatbot taking your CV and asking "tell me about yourself" five times. It is not a job board, not a resume builder, and it does not connect to LinkedIn, your email, or your calendar. The product is one session at a time and disappears with your tab.
Erase everything from this browser
Wipe your CV, the JD, the panel profiles, the brief, your answers, and the report from this browser tab. There is nothing for us to delete on the server side — there is no server side.