The AI interviewer that remembers you

You did a voice mock interview on InterviewAce on Tuesday. Zaheen — our AI interviewer — pushed back on your STAR answer because the "Result" was vague. You worked on it. You came back Friday for another session.

Tuesday's Zaheen had no idea Friday's Zaheen existed.

You got asked similar questions. You got polite generic feedback. The thing you were specifically working on? Forgotten. Every other AI interview tool on the market has this problem, and it's been the #1 reason users churn after 2-3 sessions.

As of today, that's fixed.

What we just shipped

Persistent candidate modeling. After every voice interview session, Zaheen analyzes the transcript and builds a rolling profile of you:

Weaknesses Zaheen caught (e.g. "STAR specificity", "system design tradeoffs")
Topic gaps to drill next session ("consistency models", "DP recurrences")
Hedging score — what fraction of your phrases soften claims ("kind of", "I guess", "basically")
Filler density — words per minute of "um", "like", "you know"
Pacing — your words per minute average
5-dimension rubric trend — clarity, specificity, pacing, recovery-under-pressure, depth, with signed deltas vs your prior session
One-sentence interviewer note — what the next session's interviewer should specifically probe

This model lives in your account. Every new session starts with Zaheen reading it. The first thing your next interview will do — silently, not announced — is calibrate to what you already know and what you don't.

What this fixes

Six pain points that universally drag down AI interview prep tools, all caused by the same root issue: no memory.

1. Sycophantic feedback

Every AI interview tool we tested ends with "Great answer! 9/10!" no matter what the candidate said. This is because the LLM has no context — it doesn't know you've been hedging the entire 6-minute session, or that you bombed the same topic last week. It scores each answer in isolation against a generic rubric.

With candidate modeling, Zaheen sees: "This candidate hedges in 38% of phrases. Push back when they soften claims." The scoring rubric calibrates against your historical baseline. A 65/100 on specificity used to feel arbitrary. Now it means: you were at 58 last week, 65 this week, +7 trend — keep going.

2. No cross-session learning signal

You'd practice. You'd improve. You'd have no idea you'd improved.

Now every voice-review screen shows: "Your hedging dropped 30% this session. Your clarity score improved by +4. Your recovery-under-pressure dropped by -2 — pressure-test this in your next session." The trend is the value. One session is a data point; seven sessions is a curve.

3. Same questions, every time

Without memory, the question pool is random. You get asked about hash maps in session 1 and... hash maps again in session 3, because the interviewer doesn't know you've already answered it.

With memory, Zaheen has your topic-coverage history. The system prompt now includes: "Known technical topic gaps: consistency models, DP recurrences." The next question is drawn from there, not from a random shuffle.

4. "Glorified chatbot with a mic"

Half the AI interview tools out there are essentially LLM + STT + TTS. They have voice, but they're not really interviewers. A real interviewer notices things — your tendency to over-explain on easy questions, your collapse on system design. That noticing is the entire skill.

Candidate modeling is the substrate for that noticing. Zaheen now knows that you "tend to bury impact in process — push for the metric." That's the kind of note a hiring manager writes in their head after the first round and uses to grill you harder in the second.

5. No follow-up across sessions

The classic interview prep failure mode: you practice the same things you're already good at, and avoid the things that scare you. Without external pressure, you self-select into comfortable patterns.

Now the interviewer is the one tracking what you avoided. Your topic gaps from session 1 become the opening questions of session 2. There's nowhere to hide.

6. The verbal explanation gap

The #1 reason candidates fail technical interviews isn't that they can't code — it's that they can't talk through their code. Studies consistently show this is the highest-leverage prep area, and the one most tools ignore.

Filler density and pacing tracking surface this directly. If your pacing dropped from 142 WPM to 89 WPM mid-session, that's a confidence collapse the interviewer caught — even if your final answer was "technically correct."

How it works — the technical stack

The candidate model lives as a single JSON blob in Cloudflare D1 (our SQLite-at-the-edge database), one row per user. After every voice session, the API does this:

POST /api/sessions  (your session transcript + rubric)
    ↓
Worker stores the session in D1
    ↓
Worker fires Groq Llama 3.3 70B with:
  - Your prior candidate_model (from the same D1 row)
  - The new session transcript (truncated to last 3.5KB)
  - The 5-dimension rubric scores
    ↓
Llama returns updated JSON model (~900 tokens)
    ↓
Worker writes back to D1 + returns to frontend
    ↓
Frontend caches in localStorage for instant access next session

The whole post-session update takes ~1.5-2 seconds. On the next voice session, before the first question is asked, Zaheen's system prompt receives a --- CROSS-SESSION CONTEXT --- block with your weaknesses, topic gaps, hedging tendency, and the previous interviewer's note. The next interview starts where the last one left off.

Total marginal cost: ~$0 because we're still on Groq's free tier. If/when we outgrow it, the candidate-model update costs about $0.001 per session (one 1500-token Llama 3.3 70B call). At a thousand active users running 3 sessions/month, that's $3/month total. The unit economics survive scale.

What we explicitly didn't ship (yet)

Transparent build-in-public note: this is V1. The harder version of the same feature is still on the roadmap, and we'd rather under-ship than over-promise.

V2 — Adversarial mode toggle. A new "/voice Pro+" setting where Zaheen pushes back harder mid-answer ("Wait, why O(n log n) and not O(n)?"), demands specificity on every STAR Result, and asks "what's the trade-off you're missing?" on every technical claim. The current Zaheen already pushes back, but adversarial mode would be 30% more aggressive. Coming in ~2 weeks.

V3 — True interruption via OpenAI Realtime API. Right now Zaheen is turn-based — you finish your answer, Zaheen responds. True interruption means Zaheen cuts you off mid-sentence the second you say something wrong, the way a real interviewer would. This is the highest-leverage UX upgrade but it costs $1.80/mock on the OpenAI Realtime API vs $0 on our current Groq stack. We'll ship it once we have enough Pro subscribers that the unit economics make sense — probably end of Q3.

Word-level filler density from Whisper Verbose JSON. Currently the filler-words-per-minute estimate comes from Groq Llama 3.3 70B reading the transcript and counting filler words contextually. The more precise version would use Groq Whisper's word-level timestamps to count actual filler utterances over actual elapsed time. ~1 hour of work; just hasn't been worth the dev time vs the LLM estimate which is already 90% accurate.

Try it

The feature is live right now on https://interview-prep.academy/voice (Pro plan — $10/mo, or run the free preview to see the setup flow).

The candidate model populates after your first voice session. The cross-session teardown shows up on the /voice-review screen — scroll past the per-question scores to the new section titled Cross-session candidate model.

If you want the model to update faster, do two short voice sessions (3-5 questions each) back-to-back. The diff between session 1 and session 2 is where the value is most visible.

What this means for the category

Most AI interview tools are converging on the same feature set: voice mocks, scored feedback, hire/no-hire verdict. The next 12 months will be about which tool models the candidate best, not which tool has the cleanest UI.

That's a defensible bet — models compound. After 10 sessions, your candidate profile on InterviewAce knows things about your interview style that no other tool can replicate without those 10 sessions of data. Switching costs go up linearly with usage. The flywheel is: more sessions → better model → harder & more relevant questions → better outcomes → more retention.

Marketing line, if you want it: "The only AI interviewer that remembers you."

True for now. We'll defend it.

If you've been on the fence about the $10/mo Pro tier, this is the feature that justifies it. Free tier still includes unlimited practice and mock mode — voice with cross-session memory is what the $10 unlocks.

→ Start a voice interview now · No card on free tier · Pro $10/mo with 7-day refund

Reading this in 2027? The "V2/V3 didn't ship yet" sections are probably long-done — check the changelog for the latest. The candidate-model API is at /api/me/candidate-model if you're an InterviewAce user and want to see your own raw JSON profile.