Words you will meet
- Case interview
- A consulting job interview in which the candidate solves a business problem out loud.
Chapter 1 · The problem
Mocks with classmates were hard to book, and their feedback was too general to act on
While preparing for case interviews, I could rarely book a mock with a classmate, and the feedback was general: "work on your structure". So I built a practice partner that is always there, and made it hard to start the next case before fixing one gap.
Once Case Room was running, its grader kept a summary of the gaps that recurred across sessions. Over my 7 graded sessions it found the same few gaps coming back, each in at least 3 sessions. Those recurring gaps are what the practice gate turns into steps.
Observed, 7 sessionsSources: Case Room's recurring-gap summary in its synced practice history, 28 Oct 2026.
Chapter 2 · How it works
An AI interviewer runs the case out loud, while code decides when each chart appears and when the case ends
The interviewer talks to me as on a video call: it asks questions, shares charts and pushes back. I work the math on paper and hold it up to my camera, as in a real interview.
Exhibit 1
Models do the talking and grading; code holds the rules of the case
| Part | Done by | What it does |
|---|---|---|
| Speech | Deepgram, ElevenLabs | Turn my voice into timed text, and speak the replies |
| Interviewer | Claude Sonnet | Writes each reply, ending with a short line naming the part of the case |
| Case rules | Code | Reads that line, releases a chart only after I have given a structure, and ends the case |
| Delivery | Code | Counts words a minute, filler words, long pauses and time to a structure |
| Grader | Claude Opus | Scores the session against the case's answer key and sets practice steps |
On the first day the interviewer shared a chart before I had given a structure, so that rule moved into code.
The grader scores six skills from 1 to 5, where 3 is solid and 4 is offer level. Their weighted average times 20 gives a score out of 100, so the bar of 70 means an average of 3.5, halfway between the two. To test next Chapter 5 sets out how I would compare its scores with a human interviewer's.
Sources: Case Room code and grading rubric; fix of 22 Oct 2026.
Chapter 3 · The practice gate
A new case stays locked until I pass the grader's practice steps, skip one with a reason, or carry it forward after three tries
After a session the grader writes 2 to 4 practice steps, each with a pass standard: redo one part of the case, drill that skill on a new case, or pass a 12-question math set. Until every step is done, no new case starts.
A second model marks each attempt "pass" or "not yet". Two exits keep the gate from trapping me: skip a step with a reason, which shows in red on my practice path, or carry it forward after three failed tries. I gave up the easy feeling of doing case after case, because without the lock feedback stays a note.
A Claude-written design proposal I asked for recommended a drawn interviewer face. I chose a real-looking one, on condition it adds under about 400 ms before each line, since pauses make the call less like an interview. Built Added 31 Oct. The app measures the added delay, and running that test is next.
Sources: Practice gate code; design proposal, 27 Oct 2026; commits of 31 Oct 2026.
Chapter 4 · Use so far
In its first week I ran seven sessions, and the gate set eight practice steps
Case Room has one user, me. This is everything its synced history records up to 31 October.
- 7
- sessions: 1 full case, 1 behavioral interview, 5 drills
- 1 of 8
- practice steps passed
- 3
- failed tries before one step was carried forward
Only 8 steps exist, because the gate arrived after most sessions
I added the gate on 28 October, after most sessions were graded. The app turned 5 sessions' earlier advice into one step each; 2 had none. A mock with a person, logged from a photo of its feedback sheet, added 3.
One step asked me to redo the opening of a case to a written standard. I scored 1, then 2, then 2 out of 5. After the third try the app carried it forward, so one hard step could not block practice. The step I passed was a math set: 9 of 12 on 28 October, then 11 of 12 on 31 October.
Exhibit 2
The case loop and the gate are in use; the newest parts are built and ready for their first sessions
| Part | Status |
|---|---|
| Voice case, grader, practice gate | Live 7 sessions, 8 steps |
| Library of 85 public practice cases | Built Private, behind an access code |
| Practice for three firms' online tests, and speaking practice | Built No sessions yet |
| Grader compared with a human interviewer | Next Chapter 5 |
The app estimates $0.80 of AI and speech per 30-minute case.
Sources: Synced practice history, 31 Oct 2026; commit history, 22 to 31 Oct 2026.
Chapter 5 · What comes next
The next test is to calibrate the grader against a human interviewer, then measure the new face's delay
Every practice step depends on the grader's scores, so the grader is the first thing I would test.
- Calibrate the grader. An experienced interviewer scores a few recorded sessions blind, on the same six skills. I compare the two skill by skill and adjust the rubric where they differ.
- Run the delay test. The app already measures the delay the real-looking face adds before each line. I would record the median and hold it to the 400 ms rule,.
- Check the cost. Compare the $0.80 estimate per 30-minute case with the actual AI and speech bills.
Reflections
Rules in code and a pass standard turned feedback into practice
What worked
Rules in code and a pass standard on every step. When the grader said not yet three times, the app held me to it.
What I learned
With an AI grader, I would set up a human comparison from the first session, since every practice step depends on its scores.
What I'd do next
Have an experienced interviewer score a few recorded sessions blind and compare skill by skill, then run the delay test on the new face.