Back to Projects

MockMind

[AI Coaching Interviewer]
Mar 2025 - Ongoing

An end-to-end AI Coaching Interviewer platform with real-time feedback, behavioral analysis, and personalized coaching to master technical and soft skill interviews.

Computer VisionNLPLangGraphLangChainQdrant Vector DBPython

Project Scope

What this project covers — systems owned, responsibilities, and integrations.

Multimodal AI pipeline (Voice + Computer Vision + NLP)
Real-time WebRTC voice agent architecture (LiveKit)
Post-session report generation with scoring engine
RAG pipeline for CV/JD-anchored interview questions (LangChain + Qdrant)
Full-stack monorepo — Next.js frontend, NestJS API, FastAPI AI Services

The Story

The interview process is fundamentally broken. Candidates spend months grinding LeetCode and memorizing STAR-method answers, only to freeze when speaking to a real human. To fix this, MockMind was born. MockMind is not just another ChatGPT wrapper. It's a complete multimodal feedback loop. It listens to your pacing, counts your filler words, and even watches your eye contact. It acts as a mirror that reflects your true performance, helping you build genuine confidence through rigorous, simulated confrontation.

Challenges Faced

Real technical and design problems encountered during development — and how they were resolved.

1

Latency in multi-modal real-time pipeline

Running STT → LLM → TTS in sequence added perceptible delay. The solution was parallelizing STT transcription with a VAD (Silero) pre-filter to avoid sending silence to the LLM, and caching TTS audio chunks for predictable interview question openers.

2

Consistent NLP scoring across diverse answer styles

Filler word counts and STAR-method detection were initially too rigid (regex-based). Migrated to a semantic LLM chain that reviews the full transcript in context, dramatically improving accuracy for non-native English speakers and informal speech patterns.

3

Monorepo dependency coordination (Turborepo)

Three independently deployable apps (Next.js, NestJS, FastAPI) sharing types and configs required careful Turborepo pipeline configuration to prevent stale cache issues and ensure only affected apps rebuild on code change.

Real-World Impact

Measurable outcomes and meaningful results this project delivered.

6 interview round types supported

The platform handles Technical, Behavioural, HR, Machine Coding, and fully custom user-defined interview templates.

Dual-mode feedback system

Strict Mode (real interview simulation) and Learning Mode (live coaching) serve different user needs with a single underlying AI agent, maximizing platform utility.

B2B pivot path validated

The same AI screening engine is being roadmapped as HireMind — an enterprise candidate screening product — demonstrating real commercial potential of the core architecture.

SWOT Analysis

Strengths

  • Multimodal analysis (Voice, CV, NLP) creates a holistic feedback loop
  • Personalized resource hub based on exact weakness tracking
  • Built on modern, scalable WebRTC voice agents

Weaknesses

  • High dependency on third-party LLM inference limits and costs
  • Webcam/Mic latency variations across different user hardware

Opportunities

  • B2B enterprise Pivot (HireMind) for automated candidate screening
  • Integrating peer-to-peer real human mock interview matching

Threats

  • Rapidly evolving competitor landscape in AI interview prep
  • User privacy concerns regarding storing video and voice metrics
Full Technical Documentation
GitHub