// Artifact · Full-Stack AI

PrepPilot — AI Mock Interview Tutor

A local-first AI interviewer that listens to how you speak, not just what you say — a full voice loop (STT → coaching LLM → prosody analytics → TTS) on an 8GB laptop GPU

Full-Stack AISpeech & Audio MLLocal-First LLMReal-Time Systems

Open full experience

// 01 · Introduction

Introduction

PrepPilot runs a realistic spoken mock interview and coaches the candidate on both what they say and how they say it. Unlike a chatbot, it listens to delivery — pace, pauses, filler words, pitch variance, vocal energy — and returns structured 1–10 rubric feedback with concrete rewrites, entirely offline on a laptop GPU. The demo video below shows the full voice loop; the complete source and system design are on GitHub.

// 02 · Description

Description

A candidate picks a role, seniority, and optionally pastes a job description. An AI interviewer asks one question at a time over a WebSocket voice loop: Silero VAD detects the end of an answer, faster-whisper large-v3 transcribes it with word-level timestamps, a prosody/filler analytics layer computes interpretable delivery metrics (WPM, pause ratio, filler rate, documented confidence and expressiveness composites — deliberately not black-box emotion labels), and an LLM scores the answer on a five-part rubric (content, structure/STAR, specificity, technical accuracy, delivery), choosing between a targeted follow-up and the next question. A Kokoro TTS voice asks it aloud. Sessions persist to SQLite, feeding a printable report and a cross-session trends dashboard. The LLM is swappable — local Ollama (qwen3) by default, Claude or GPT with one config line — and every heavy dependency degrades gracefully: the full text-mode loop works with zero ML packages installed.

// 03 · Objective

Objective

To demonstrate the ability to design and ship a complete, production-shaped AI system — not a notebook or a single prompt — that integrates real-time audio, multiple ML models, and an LLM under a hard resource budget (8GB VRAM), while making deliberate engineering trade-offs around reliability, privacy, and cost.

// 04 · Process

Process

01Designed the contract first: shared Pydantic schemas and a documented WebSocket protocol between the FastAPI backend and the TypeScript frontend, so every later feature had a stable spine.

02Built an async turn state machine (orchestrator) that is transport- and provider-agnostic — the live voice WebSocket path and the REST endpoints drive the same logic, and interviews survive a server restart mid-session.

03Assembled the audio pipeline: Silero VAD end-of-turn detection (with an energy fallback and a manual flush so a quiet mic never silently drops an answer) → faster-whisper transcription with word timestamps → parselmouth/librosa prosody analysis → Kokoro TTS reply.

04Made the LLM swappable behind a provider abstraction (Ollama, Anthropic, OpenAI, and a canned offline provider) with automatic fallback; all model output is parsed as strict JSON with a retry-on-validation-failure loop.

05Engineered the 8GB VRAM budget explicitly: two model-residency strategies (a 4B LLM co-resident with Whisper, or an 8B LLM loaded on demand between turns) documented and configurable.

06Enforced graceful degradation as a hard rule — every heavy dependency imports lazily, a health endpoint reports machine-readable degradation flags, and the UI surfaces them honestly instead of failing silently.

07Verified with 58 GPU-free unit tests plus end-to-end voice testing on the target hardware, then published the source, a system-design document, and the recorded demo.

// 05 · Tools & Technologies

Tools & Technologies

Python · FastAPI · WebSocketsfaster-whisper large-v3 (STT)Silero VADKokoro-82M TTSparselmouth + librosa (prosody analytics)Ollama (qwen3) · Claude · GPT (swappable providers)SQLite · SQLAlchemy 2.0Next.js 16 · TypeScript (static export)Docker · pytest · CUDA (RTX 3070, 8GB)

// 06 · Value Proposition

Value Proposition

Unique Value

Most AI interview tools grade the transcript. PrepPilot treats delivery as a first-class signal using interpretable, documented metrics rather than black-box emotion labels — and it does this local-first, so the candidate's voice never leaves their machine. Fitting real-time STT, an 8B-parameter LLM, and TTS inside an 8GB VRAM budget, with graceful CPU-only degradation and reconnectable sessions, is the core engineering achievement: it is systems design, not prompt-writing.

Relevance to My Audience

Interviewing under pressure is a near-universal professional bottleneck, and delivery is where strong candidates most often lose points — PrepPilot gives job seekers a private, unlimited, judgment-free way to rehearse. For hiring managers, it is end-to-end evidence of the skills an AI/ML engineer is actually hired for: real-time inference, model orchestration, resource budgeting, provider abstraction, and reliable UX under failure.

// 07 · References