IntervAI

Hyper-personalized, face-to-face AI mock interviews powered by Google Gemini—tailored to your resume and target job description.

Comment

Inspiration

Job interviews are stressful, but traditional preparation is fundamentally broken. Candidates spend hours reading static question banks or practicing answers in front of a mirror. Neither approach replicates the real pressure of a face-to-face conversation: dynamic follow-up questions, evaluating vocal confidence under stress, or probing specific claims on a resume.

We built IntervAI to bridge this gap. We wanted to build a realistic, zero-judgment, face-to-face AI interviewer that knows your resume and your target job description inside out—giving candidates a safe space to fail, learn, and build real confidence before high-stakes interviews.

What it does

IntervAI is an interactive, voice-and-avatar AI mock interview platform that transforms job preparation into an immersive practice environment:

  1. Resume & Job Description Grounding: Candidates upload their resume and target JD. The system parses requirements, extracts key competency benchmarks, and crafts a custom interview plan.
  2. Real-Time Face-to-Face Avatar: Candidates speak directly with a photorealistic video avatar that listens, responds, and conducts the interview naturally in real time.
  3. Adaptive Difficulty Engine: Questions adapt dynamically based on candidate responses—diving deeper into technical details if the candidate excels, or pivoting to core concepts if they struggle.
  4. Comprehensive Performance Analytics: Post-interview, IntervAI generates a detailed report featuring an overall Readiness Score, section breakdowns, model answer suggestions, and Resume Gap Detection (flagging claims on the resume that weren't backed up live).

How we built it

  • Frontend & UI: Built with Next.js 14, React, Tailwind CSS, and Framer Motion for a smooth, high-performance interface and live video overlay.
  • Backend Infrastructure: Powered by FastAPI (Python) with MongoDB handling asynchronous session state and report storage.
  • Intelligence & Grounding: Google Gemini API (gemini-2.5-flash) drives resume parsing, job requirement extraction, dynamic question generation, and report evaluation.
  • Voice & Avatar Pipeline: Integrated Tavus CVI for low-latency interactive video avatars, combined with Groq Whisper for rapid speech-to-text (STT) transcription.
  • Scoring & Adaptive Math Engine: Overall candidate readiness $S_{\text{readiness}}$ is calculated using a weighted multi-criteria formula:

$$S_{\text{readiness}} = 0.35 \cdot s_{\text{tech}} + 0.25 \cdot s_{\text{fluency}} + 0.20 \cdot s_{\text{tone}} + 0.20 \cdot s_{\text{relevance}}$$

Difficulty levels adjust dynamically turn-by-turn according to candidate performance:

$$D_{t+1} = \text{clamp}\left( D_t + \delta \cdot \text{sign}\left(s_{\text{turn}} - \theta\right), \, 1, \, 5 \right)$$

where $D_t \in [1, 5]$ represents difficulty, $s_{\text{turn}}$ is the turn score, and $\theta = 70.0$ is the performance threshold.

Challenges we ran into

  • End-to-End Latency Optimization: Maintaining conversational flow required keeping overall round-trip latency $\tau_{\text{total}} \le 1.2\text{s}$. We optimized our backend pipeline by streaming Groq Whisper transcriptions directly into Google Gemini API prompts.
  • Contextual Grounding vs. Hallucinations: Ensuring the AI interviewer strictly strictly adhered to the candidate's uploaded documents required carefully crafted system prompts and structured JSON response schemas.
  • Resume Gap Verification: Designing a prompt heuristic that fairly evaluates whether a candidate's spoken response satisfies a claimed skill on their resume without penalizing natural verbal phrasing.

Accomplishments that we're proud of

  • True Face-to-Face Interaction: Moving beyond text-based AI bots to deliver a natural voice-and-video avatar experience.
  • Hyper-Personalized Content: Ensuring no two interviews are ever identical—every question stems directly from the user's background and target role.
  • Actionable Post-Interview Diagnostics: Delivering structured feedback, score metrics, and model answers that give candidates clear steps to improve.

What we learned

  • How to leverage Google Gemini API for structured JSON extraction and evaluation tasks.
  • Techniques for handling real-time audio/video streaming, WebRTC events, and asynchronous turn management in FastAPI.
  • The psychological value of avatar-based practice: realistic vocal and visual feedback significantly reduces candidate anxiety compared to text-only tools.

What's next for IntervAI

  • Interviewer Personality Selection: Allowing candidates to choose interviewer personas (e.g., rigorous technical lead, supportive HR manager, fast-paced founder).
  • Interactive Code & Whiteboard Sandbox: Adding live coding environments for technical software engineering rounds directly inside the interview view.
  • Longitudinal Skill Tracking: Tracking candidate performance trends over multiple practice sessions to measure progress over time.

Built With

Share this project:

Updates

Submission history