AI Interview Simulator

I built this after a job interview where I never spoke to a person. Being interviewed by a bot was deeply frustrating — it was slow to answer, quick to talk over me, and cut my answers short before I could finish. It is the kind of AI-led first interview people joke about online, until it happens to you.

I had just been interviewed by an AI. So I built a better one to practice with.

Every frustration from that call became a design choice — I wanted mine to be everything that bot was not. It waits until you have truly finished before it replies, and answers one sentence at a time, so a turn feels like a real conversation instead of a fight to be heard.

You give it your CV and a job description, then talk through a real voice interview. Four parts work together on every turn:

• Listen — your speech becomes text with Faster-Whisper.
• Lead — a LangGraph agent asks questions matched to your CV and the job (searched with Qdrant).
• Speak — its replies are spoken back to you with Piper.
• Score — when you finish, a second AI writes a report that links each point to something you actually said, not just a number.

FastAPI backend (data in SQLModel + Alembic), React + TanStack Start front end, with the AI models running locally on Ollama. Docker Compose runs the API, database, vector store, and speech services together.

One rule holds it together: every part talks through a clear interface. Speech-to-text, text-to-speech, and turn detection are used only through STTProvider, TTSProvider, and TurnDetector, and a setting picks which one. Moving the model, vector store, or database to the cloud is a config change — not a rewrite.

How a single turn works

Ollama native, not in Docker. On Apple Silicon, Docker cannot use the GPU, so Ollama in Docker runs CPU-only and crawls. Running it natively uses the Apple GPU (Metal) and is far faster. Cost: no single docker compose up — but for real use, the speed is worth it.

Turn detection: a simple rule, not a big model. By default the app uses a light rule over the words so far — no heavy model, no restrictive license bundled in. You can opt into a stronger model instead; I documented the trade-off for each:

They take different inputs — some audio, some text — so the TurnDetector interface carries both. Any of them drops in without touching the rest of the app.

Voice is an optional extra. It installs on demand (uv sync --extra voice) and loads only when needed, so the text-only version and the tests stay small and fast. Cost: voice users run one extra install step.

Because the AI runs locally, no data leaves your computer and there is no per-request bill. The same code runs on a laptop or in the cloud, with no change to the app logic.

And I measure it instead of guessing. The app times each step — turn_detection_ms, stt_ms, graph_ms, tts_ms, plus time-to-first-audio — sends them to LangSmith, and warns me when a step runs slow. On a local GPU the interviewer replies about 1–3 seconds after you stop (≈0.7–1s in the cloud).

• Clear interfaces from the start let the app run both locally and in the cloud without writing the code twice.
• How fast it feels beats the total time.* Speaking back one sentence at a time helped more than any other speed fix.
• Measure before optimizing. Timing each step in LangSmith showed the real slow point, instead of guessing.