Skip to content
← All products

Personal build · Oct 2026 · Built for my own use

Case Room, a voice case-interview simulator I built for my own practice

A voice simulator for structured problem solving out loud under questioning, in the form of a case interview. An AI interviewer runs the case, a second model grades it, and the next case waits until I do the practice it set.

At a glance

Role
Built it alone for my own practice: the product, the prompts and the code, from 22 to 31 Oct 2026.
Team
Just me. Claude Code agents wrote much of the code; I set the design and reviewed every change.
Key decision
Lock the next case until the grader's practice steps are done, with two exits: skip with a reason, or carry forward after three failed tries.
Result
Built and in use: 7 sessions in the first week. 1 of 8 practice steps passed; 1 carried forward after three failed tries.
What I'd change
Calibrate an AI grader against a human interviewer from the first session, with a few sessions scored blind, skill by skill.

TL;DR

Problem
Mocks with classmates are hard to book, and their feedback is general, such as "work on your structure".
Built
A voice interviewer, a grader that scores six skills, delivery counts, 85 public practice cases, and a gate that holds the next case until the practice is done.
Evidence
7 sessions on 27 and 28 Oct 2026. Of 8 practice steps, 1 passed, 1 carried forward after three failed tries, 6 open on 31 Oct.
Next
Have an experienced interviewer score recorded sessions blind to calibrate the grader, then run the delay test on the new interviewer face.

Key numbers

1 of 8
practice steps passed; 1 carried forward after three failed tries, 6 still open
Synced practice history as of 31 Oct 2026.
85
public practice cases in the app's library
Built, private, behind an access code. Commit history, Oct 2026.
$0.80
estimated AI and speech cost of one 30-minute voice case
The app's own cost estimate (lib/prompts.js).

The problem

Case practice needs feedback specific enough to act on

A case interview asks you to solve an unfamiliar business problem out loud: structure it, work the numbers, read a chart and land a recommendation while someone questions each step. Mocks with classmates were hard to book, and their feedback was general, such as "work on your structure". So I built a practice partner that is always there, turns its feedback into practice steps, and holds the next case until I do them.

Key decisions

Code keeps the rules; the next case waits

Decision 01

Put the rules of the case in code, not the prompt

The call
The interviewer model ends each reply with a short line naming the part of the case. Code reads it, releases a chart only after I give a structure, and ends the case.
What I gave up
A simpler build in which the model runs the whole case.
Why
On day one the interviewer shared a chart before I had a structure. A rule the model can forget is not a rule.
Decision 02

Lock the next case until the practice is done

The call
The grader sets 2 to 4 practice steps, each with a pass standard, and a second model marks each try. No new case starts until every step is done.
What I gave up
The easy feeling of doing case after case.
Why
Without the lock, feedback stays a note. Two exits stop it trapping me: skip with a reason, shown in red, or carry forward after three failed tries.

What I built

A voice interviewer, a separate grader, delivery counts

The interviewer, Claude Sonnet, talks to me as on a video call and pushes back; I hold my paper math up to the camera. Code counts words a minute, filler words, pauses and time to a structure. Claude Opus grades six skills from 1 to 5 against the case's answer key. The library holds 85 public practice cases.

Screenshot to addThe practice path after a graded session: the six skill scores, the practice steps with their pass standards, and one step skipped in red. Annotate the lock on the next case. Show no case content.

I chose a real-looking interviewer face over the drawn one a design proposal suggested, on condition it adds under about 400 milliseconds before each line. The app measures that delay, and running the test is next.

Use so far

In week one, the gate held me to it

I am the only user so far.

WhatCount, to 31 Oct 2026Status
Sessions (1 full case, 1 behavioral, 5 drills)7Live
Practice steps passed1 of 8Observed
Carried forward after three failed tries1 of 8Observed
Grader compared with a human interviewerPlannedNext

I passed a 12-question math set, 9 of 12 then 11 of 12. A redo of a case opening scored 1, 2, then 2 out of 5. After the third try the app carried it forward, so one hard step could not block practice.

To fill Real usage: sessions, people, how many came back in a second week

Next

The next test: calibrate the grader against a human

Every practice step depends on the grader's scores, so it is the first thing I would test. An experienced interviewer would score a few recorded sessions blind, and I would compare the two skill by skill. Then I would run the delay test on the new interviewer face against the 400 millisecond rule, and check the $0.80 estimate against the actual bills.

Sources: Case Room code, rubric and commit history, 22 to 31 Oct 2026; synced practice history, 31 Oct 2026; design proposal, 27 Oct 2026.