Skip to the work
← All work

AI product

Hackcepted

Rehearse the interview before it counts.

Live in productionhackcepted.com →
TypeScriptReactNext.jsVercelPostgresStripeReal-time avatars

The question it started from

What if you could sit through the interview you are dreading as many times as you needed, and be told exactly where you lost the room?

The problem

Interview preparation is advice, not practice. You read lists of questions, rehearse in your head, and then face a real person for the first time on the day it matters most.

The gap is not knowledge. It is that nobody gets to see themselves do it badly first.

The idea

Put a believable interviewer in a real video call, let it follow up the way an interviewer does instead of reading a script, and turn the recording into feedback that points at specific moments.

What I contributed

I built the product: the AI interviewer that joins a real video meeting, the scheduling and launch flow around a session, the analysis that turns a conversation into a report, and the accounts, credits and payment layer.

I also built the landing and the acquisition side, and ran the experiments that told me which promises people actually responded to.

A session
  1. 01

    Pick the role

    The interview is set up around the job you are actually going for.

  2. 02

    Join the call

    The AI interviewer joins a real video meeting as a participant.

  3. 03

    Be interviewed

    It follows up on your answers instead of reading down a list.

  4. 04

    Analyse

    The session is turned into structured feedback rather than a score.

  5. 05

    Report

    You get the specific moments where it went wrong, and why.

What exists today

  • An AI interviewer with a face and a voice that joins a real video meeting as a participant.
  • Distinct interviewer personas, each with its own voice and opening.
  • A session flow: pick the role, get the call, do the interview, receive the analysis.
  • Accounts, credits and checkout around the whole thing.
  • All of it deployed and open at hackcepted.com — landing and product out of the same codebase.

What is left to build

  • The report trigger needs to fire reliably on its own rather than depending on a single webhook.
  • The subscription rebuild is built but not yet in front of users.

What it taught me

Latency is a product feature. An avatar that felt dead on arrival was not a model problem — the fix was to speak the greeting immediately, before any model round-trip, so the first second of the call felt alive.

Identity plumbing deserves as much care as the prompt. When two personas shared one conferencing identity, the wrong interviewer silently joined the call: a bug no amount of prompt tuning would have found.

Clicks are not intent. An acquisition campaign that performed well on clicks produced nobody who wanted the product, which taught me to judge an experiment by behaviour further down rather than by the metric the platform optimises.

Next

Dotly