Guided interview practice rooted in questions, speech, and reason
An AI-powered behavioral interview coach. Speak your answer, get a tailored follow-up question and rubric-based scoring in the same flow you'd get from a real interviewer — including how you came across on camera.
Now in beta with college undergraduates, new grads, and recruiters / hiring managers helping calibrate the coaching against the bar real interviewers set.
Behavioral interviews decide who gets the offer, but they're the part of the loop candidates rehearse the least and lose offers on the most. Friends can't simulate a stranger pushing back with a sharp follow-up, mock-interview platforms skew technical, and recording yourself gives you a tape without coaching — you hear the rambling but not which part is hurting you.
InterviewPie closes that gap: speak your answer to a real interviewer voice, get a follow-up question that references what you actually said, and receive six-dimension scoring grounded in your industry plus a delivery grade pulled from your webcam. It's built for college undergraduates preparing for their first internship loops, new grads navigating full-time hiring, and the recruiters and managers who care that what candidates are practicing actually maps to what gets people hired.
InterviewPie classifies every session into one of 15 field/industry buckets (Tech, Finance, Healthcare, Legal, Consulting, Sales, Ops, Nonprofit, Education, Government, and more) using both the company and target job title, so cross-functional roles land in the right place — a healthcare counsel role is graded as Legal, not as Healthcare. That classification drives two downstream choices: the opening question is written by a system prompt tailored to the field's actual interview shape, and the evaluator is handed a rubric whose criteria match that field. A finance candidate's "Problem Solving" is judged on quantitative trade-offs; a healthcare candidate's is judged on patient-safety reasoning. The result feels less like a generic STAR drill and more like preparing for the screen you're actually walking into.
The first question of every session is shaped by live research, not pulled from a static bank. When a session starts, InterviewPie fires two parallel Google searches via Serper — one for the company in general, and one specifically for its behavioral interview style and culture for the candidate's target role (the generic "interview questions" corpus was deliberately dropped because it's dominated by LeetCode and system-design content). A Gemini summarization pass distills both into a compact brief that includes two new fields the opening-question generator conditions on: role values — what the company is documented to look for in applicants for this specific role (cultural and soft-skill traits only — technical proficiencies are explicitly excluded), and common question themes — short labels (never verbatim questions) drawn from any behavioral interview leaks the search surfaced. For small or obscure companies where no concrete signal exists, both fields come back empty rather than invented, and the question generator falls back cleanly to the field-tailored defaults instead of hallucinating a role profile.
On top of the brief, the generator samples two of five example question shapes per call from a per-field-category pool. That randomized rotation breaks the "fixed attractor" effect where the same model running the same prompt converged on the same question across companies — a beta-tester complaint that "Tell me about a time you faced a difficult technical challenge…" arrived verbatim at five different big-tech screens in a row. The post-session summary surfaces the same role values and question themes as bullets in the Company brief card, so the candidate can see what shaped the questions they were just asked.
To stop the same opening question recurring across back-to-back sessions, InterviewPie caches each candidate's three most recent opening questions and feeds them back into the generator as an explicit avoid-list, so the next question has to take a distinct angle. The cache is scoped to the candidate (not the company), since opening questions track role and seniority more than which company you're interviewing at, and it resets whenever you change your target role, industry, or experience level — the point at which older questions stop being relevant.
Every answer is scored 0–10 on six dimensions: Structure, Problem Solving, Impact, Initiative, Depth, and Delivery. Five come from the LLM evaluator; Delivery comes from your webcam (see next section). Each turn ships back with structured coaching that balances what worked with what to fix. The evaluator quotes exact transcript snippets for 1–3 positive moments ("keep doing this") and 2–4 improvement moments, then gives bite-sized suggestions such as adding the customer's actual concern, one reasoning sentence, or a small outcome. The goal is specific feedback without turning the product into a full answer generator.
Every answer is scanned for the usual verbal crutches — um, uh, like, you know, basically, actually — but a raw tally is misleading: a longer, better answer almost always contains more of them. So InterviewPie reports the filler-word rate: the percentage of your words that were fillers. A four-minute answer with twelve "likes" can be cleaner than a thirty-second one with four, and the rate is what tells you that. It shows up as a color-graded bar — green under 5%, yellow to 10%, orange to 15%, red above — on the session overview, under each turn's scores, and as a "filler rate over time" trend on your History page so you can watch it fall across sessions. The lifetime figure is word-weighted, so one rambling session can't quietly skew it, and the History page also breaks out your five most-used filler words so you know which habit to target first.
A 478-point MediaPipe face-landmark mesh runs in the browser at 15 fps while you're answering, tracking eye contact, gaze stability, head pose, expression, and face visibility. Those signals roll into the Delivery score and into the structured feedback — so you'll get pointed feedback like "Eye contact landed at 47/100; pick a spot near the camera and return to it between phrases." This is the part of InterviewPie we lean on hardest: rather than a single averaged number, the Delivery score is computed deterministically from those signals, weighting the cues recruiters actually react to — eye contact and staying on-camera — far above softer ones, so you're never docked for a calm, even expression the way you are for staring at the floor or dropping out of frame. The feedback then breaks out into separate plain-English cues for eye contact, framing, posture, and expression, each with its own score and fix, so you know exactly which one is costing you and what to do about it. On the replay screen, the face-mask overlay redraws the landmarks on your recording so you can see what the model saw. Decline the camera and everything else still works.
Onboarding takes a PDF resume and a short bio, extracts the text, and captures industry, target role, and experience level. Every downstream prompt — the opening question, the follow-up, the evaluator — is conditioned on that profile, so the interviewer references your actual projects, internships, and seniority instead of asking a stock question about teamwork.
Experience level is a first-class second axis on top of the 15 field buckets: the company research, opening question, and evaluator rubric are all re-tuned to your seniority, so an intern is probed on learning-in-ambiguity and scored on coachability while an executive is probed on portfolio bets and scored on enterprise leadership — same company, very different interview.
The whole session runs through voice: question audio plays, the mic engages automatically when it ends, you talk, you press End answer, and the next question arrives. You choose the session length up front — anywhere from 2 to 8 turns — and the interview is broken into short "story blocks": a fresh opening question followed by one or two follow-ups that drill into what you actually said. Whether a second follow-up comes is decided on the fly — a lightweight model call judges whether your answer left a gap worth probing, so a complete answer moves on to a new scenario instead of being padded with a forced extra question. Each follow-up is conditioned on the same field category and company-research signals (role values, behavioral themes) that shaped the opening question, so it lands in the right tone for the role and gently redirects rather than echoing back if your answer was off-topic or nonsensical. You can pick from a pool of accented interviewer voices (or let the system surprise you) so non-native English speakers can rehearse against the kind of voice they'll actually face in a screen. The chosen voice persists across every turn so the interviewer never "changes person" mid-session.
If you have to stop early, the turns you already finished aren't thrown away: quitting a session with at least one completed answer scores and saves it on just those turns, so an interrupted interview becomes a shorter graded one rather than a total loss.
Every completed session is persisted with its transcript, scores, audio replay, and filler-word breakdown. The History page renders an interactive trend chart of all six dimensions across every session you've ever done — so improvement (or regression) on Structure, Impact, or Delivery is visible at a glance instead of guessed at. You can open any past session and re-listen to your own answer next to the score that explains why.
Most prep tools are one-and-done: you answer a question once and never see it again. InterviewPie lets you save an opening question straight from the results screen and come back to it later, so you can drill the same scenario until the answer is sharp. Each saved question keeps its own progress chart — every re-practice is one point, oldest to newest, plotting the opening answer's per-dimension scores and their average, so you can watch a specific question get better instead of inferring it from your overall trend.
The key design choice is that a saved question is a frozen snapshot, not a live pointer. The company, the job title, and your experience level are captured the moment you save, and every re-practice replays that exact setup even if you later change your profile — so the follow-up framing and the evaluator's rubric stay identical across attempts and the comparison is genuinely apples-to-apples. (Fixing this also closed a latent bug where changing your experience level mid-session could shift the second turn's rubric.) Re-practice is fast because it skips the company-research and question-generation model calls entirely and reads the frozen question back; only the interviewer voice is regenerated, so you can hear it in a different accent each time. You can keep up to five saved questions at once, and deleting one never erases the practice sessions behind it — they stay in your history.
InterviewPie currently ships a single Free tier, capped at 5 completed interview sessions per day. The counter resets at midnight in your own local timezone (not server time), and only ticks up when a session actually finishes — abandoning mid-session doesn't burn a slot. A Pro tier with unmetered sessions is on the roadmap but not yet exposed; everyone is on Free today.
The backend keeps an internal incidents table for security and product
auditing. It records backend-observed account creation/sign-in events,
OpenAI moderation requests, successful interview session starts, and server
errors. Incident writes are best-effort and isolated from the main request so
the audit trail never breaks an interview flow; moderation payloads and errors
are capped before storage. This log is intended for admins and debugging, not
for regular user-facing APIs.
A small, deliberately boring stack — React + FastAPI + Postgres, with a handful of sequential LLM calls per session that scale with its length (no multi-agent loop). The frontend is React 19 + Vite + Tailwind 4 with Clerk for auth, MediaRecorder for capture, and MediaPipe Tasks Vision for the in-browser face landmark mesh. The backend is FastAPI on async SQLAlchemy with Alembic migrations against Postgres, all LLM calls routed through OpenRouter. The AI layer uses Google Gemini 2.5 Flash for company research, field classification, and question generation; DeepSeek v3.2 for the evaluator; ElevenLabs for both speech-to-text and text-to-speech; and Serper for the Google search that grounds the company brief.
Browser (React + Vite)
│── Clerk JWT ──────────────────────► FastAPI
│── MediaRecorder blob (audio) ─────► FastAPI ── ElevenLabs STT
│── cv_summary JSON sidecar ────────► FastAPI
│ │── Serper + Gemini 2.5 Flash (research + field classification)
│ │── Gemini 2.5 Flash (opening + follow-up question)
│ │── DeepSeek v3.2 (evaluator, field-tailored rubric)
│ │── ElevenLabs TTS (interviewer voice)
│ └── Postgres (sessions, turns, metrics, saved questions, incidents)
└── base64 audio data URL ◄──────────── FastAPI
- Spoken-feedback mode — pipe the structured feedback back through TTS at the end of the session so the review feels like a debrief, not a report card.
- Comparative analytics — anonymized cohort percentiles ("your Structure and Impact scores trail the median for entry-level SWE candidates") would turn the trend chart from a self-comparison into a benchmark.