INSPIRATION
I built InterviewOS because interview preparation is broken. Whether you're preparing for a tech role at FAANG, a finance position at a top fund, or a law position at a prestigious firm, the feedback loop is fundamentally flawed. Most people either practice alone (zero feedback) or wait weeks for actual interview results to learn what went wrong.
I realized that Gemini 3's advanced reasoning capabilities could solve this instantly. It's not just a model that can score answers—it can explain why an answer works or doesn't work, pinpoint exactly what to improve, and generate harder follow-up questions to deepen understanding. That's real coaching, not just evaluation.
The insight: If we democratize interview coaching with AI, we level the playing field. Right now, people with mentors or money for expensive coaches have an unfair advantage. InterviewOS changes that.
WHAT IT DOES
InterviewOS is a web application where you:
- Select your domain: Tech, Finance, or Law
- Choose your mode: Structured Learning Path (4 progressive levels) or Random Practice (any question)
- Get a question: Domain-specific, realistic interview questions
- Answer: Type or speak your response using your microphone (Web Speech API)
- Get instant feedback: Gemini 3 evaluates your answer in real-time on Structure, Clarity/Confidence, and Technical Accuracy
- Learn and improve: See your scores, read personalized coaching tips, review specific improvements, and tackle a follow-up question
- Track progress: Earn badges, unlock new levels, watch your skills improve
HOW WE BUILT IT
I chose a web-first tech stack optimized for speed and user experience:
Frontend: React with TypeScript for type safety and developer experience Styling: Tailwind CSS with MIMO design system (purple #6C5CE7, orange #FF9F43, teal #00B894) AI Engine: Google Gemini 3 API for intelligent reasoning and feedback Backend: Firebase (Authentication for progress tracking, Firestore for storing user data and scores) Voice: Web Speech API (native browser speech recognition) Deployment: Firebase Hosting for instant global deployment
Architecture Philosophy: Keep it simple and fast. No unnecessary microservices, no complex backend logic. Let Gemini 3 do the heavy lifting, use Firebase for data persistence, and focus on creating an exceptional user experience.
Key Technical Decisions:
Web-First Approach: I chose web over native because it's faster to build, easier for judges to test (no app store approval), and reaches more people instantly.
Text and Voice Hybrid: I implemented both text input and voice input using Web Speech API. This serves different user preferences: text for introverts and people who think better while typing, voice for authentic interview practice.
Gemini 3 Prompt Engineering: The core innovation is the prompt design. I crafted domain-specific prompts that request structured JSON responses, specify exact scoring criteria, and ask for strengths, improvements, coaching tips, and follow-up questions in one call.
Learning Path Progression System: Instead of just a random question generator, I built a 4-level learning path for the Tech domain. Users unlock levels by completing previous levels with a 6.0+ average score. This creates natural progression and prevents jumping to questions they're not ready for.
Achievement Badges: To drive engagement and make progress visible, I added badges like Fundamentals Master (avg 8.0+ on Level 1), Algorithm Solver, System Architect, Design Expert, Streak Master, and Perfect Score.
MIMO Design System: I studied the MIMO app and adopted their design philosophy: bright, friendly colors (not corporate blue), generous spacing and padding, 16px border radius on everything, encouraging warm copy, and clear visual hierarchy.
CHALLENGES WE RAN INTO
Gemini 3 JSON Parsing: Gemini 3 sometimes returns JSON wrapped in markdown backticks. I implemented regex-based JSON extraction to handle both formatted and unformatted responses gracefully.
Web Speech API Browser Compatibility: Web Speech API is not standardized across browsers. Chrome and Edge have full support, Safari has partial support, Firefox has limited support. I implemented feature detection and provided fallback to text input if voice is unavailable. I also showed helpful error messages: "Your browser doesn't support voice input. Try typing instead."
Firebase Security Rules: I needed to allow anonymous users to practice, authenticated users to save progress, questions to be publicly readable, and user progress to be private. I carefully configured Firestore rules to allow public read on questions but private read/write on user progress.
State Management Complexity: Tracking user progress across multiple levels, questions, scores, and badges gets complex quickly. Gemini API calls are async, voice recording is async, Firebase reads and writes are async. I solved this by using React hooks for local state, custom hooks like useGeminiEval and useVoiceRecognition, and async/await with proper error handling.
Voice Transcription Accuracy: Users speak in different accents, speeds, and environments. Web Speech API sometimes misheard words or missed phrases. I solved this by letting users see and edit the transcript before submitting, providing a re-record button, showing interim results as they speak, and auto-stopping after 5 minutes of continuous speech.
Performance: Gemini 3 API Latency: Gemini 3 API calls take 2-5 seconds on average. I showed loading states with progress messages ("Analyzing your answer with Gemini 3..."), added spinner animations, cached responses, and added estimated time messages.
Question Quality and Realism: Generating high-quality, realistic interview questions algorithmically is hard. I hand-curated question banks instead of generating them. I included 20 tech questions, 15 finance, and 15 law questions based on actual interview questions from top companies.
Time Management (17-Hour Hackathon Constraint): I ruthlessly prioritized. I dropped Expo (web is enough), video recording (voice and text cover 95% of use cases), team features, and admin dashboards. I focused on the core loop done really well: question to answer to feedback.
ACCOMPLISHMENTS THAT WE'RE PROUD OF
Full-Featured MVP in Hours: The app is genuinely functional end-to-end. Select domain, get question, answer (text or voice), get Gemini feedback, and track progress. Learning path with 4 levels, unlock logic, and progression. Voice input with real-time transcription. User authentication and progress persistence. All deployed and publicly accessible.
Smart Gemini 3 Integration: The feedback isn't generic. It catches real issues (rambling, vagueness, overconfidence, missing details), provides specific and actionable improvements, generates contextually relevant follow-up questions, adapts scoring to the domain, and explains why an answer works or doesn't.
Excellent User Experience: Mobile-first design that works flawlessly on phones. Voice input with real-time transcription and edit capability. Visual feedback with scores and progress bars color-coded (green/orange/red). Learning path with clear progression, unlock logic, and badges. Encouraging tone throughout.
Learning Path as Differentiator: Most interview apps are just random question generators. InterviewOS has a structured 4-level learning path that starts with fundamentals (builds confidence), progresses logically (data structures to system design to advanced design), unlocks based on performance, awards badges for motivation, and makes users feel like they're on a learning journey.
Accessibility and Inclusivity: Text input for people who prefer typing or have speech difficulties. Voice input for authentic interview practice. Clean, high-contrast design (WCAG 2.1 AA compliant). Works on any browser with fallbacks. Anonymous mode (no account required) plus optional login for progress saving.
Production-Ready Code: TypeScript throughout (type safety). Proper error handling (network errors, API failures, permission denials). Loading states on all async operations. Clean component structure. Firebase security rules. Environment variables for sensitive keys. SEO optimization. Lighthouse 90+ performance score.
Designed for Viral Sharing: The score-sharing moment is built in. You get 8.2/10 on a tech question and want to tell your friends. Social sharing buttons on results. Badges create FOMO.
WHAT WE LEARNED
Reasoning Models Change Everything: Gemini 3's ability to explain why is what makes this app useful. A simple classifier model that just scored 1-10 wouldn't work. People need to understand what went wrong and how to improve. Gemini 3's reasoning capabilities let us provide that. Key insight: When building with LLMs, focus on what makes reasoning valuable.
Constraints Drive Innovation: With only 17 hours, I couldn't build everything. But cutting features forced better prioritization. Web beats Expo (faster to iterate). Text and voice beats video (covers most cases, easier to implement). Learning path beats random practice (higher engagement). Curated questions beat generated questions (better quality). Key insight: MVP doesn't mean minimal features. It means maximum focus on core value prop.
UX Matters More Than Features: I spent as much time on encouraging copy, loading state messaging, progress visualization, and error messages as I did on core features. And it was worth it. A beautiful, encouraging, intuitive app with 5 features beats a clunky app with 10 features. Key insight: People don't use features. They use experiences.
Prompt Engineering is a Skill: Getting Gemini 3 to return exactly the JSON format needed with high-quality coaching feedback took iteration. Be explicit about responses, specify format, give examples, and use domain-specific prompts. Key insight: Prompting is engineering.
Authentication and Data Persistence is Important: I realized halfway through that persisting user progress to Firebase was crucial. Users need to see improvement over time. Badges motivate when you can see progress. Learning paths only work if progress is saved. Without accounts, the app is a toy. With accounts, it's a tool.
Web Speech API is Surprisingly Good: I was skeptical about Web Speech API but it's fast (results in less than 1 second), accurate (95%+ accuracy in quiet environments), free (no API calls needed), and cross-platform. The only caveat: needs HTTPS and user permission. Key insight: Sometimes the native browser API is better than adding a third-party library.
Design Systems Matter: I adopted MIMO's design system (colors, border radius, spacing) and it made everything faster to build (no design decisions, just follow the system), more cohesive (consistent throughout), more polished (pre-designed patterns look professional), and easier to maintain. Key insight: Copy an existing design system for MVP. Don't try to design from scratch during a hackathon.
WHAT'S NEXT FOR INTERVIEWOS
Short-term (If This Gains Traction):
- Add Finance and Law learning paths (currently only Tech is structured, Finance and Law are random)
- Progress tracking dashboard (see your improvement over time)
- Leaderboards (friendly competition with friends)
- More questions (expand each domain from 15-20 to 50+ questions)
- Integration with LinkedIn (auto-populate role-specific questions)
Medium-term:
- Video recording option (judge body language and speech)
- Peer feedback mode (practice with friends, get human feedback)
- Industry-specific question banks (FAANG style vs startup vs enterprise)
- Accessibility features (captions for deaf users, tone analysis for non-native speakers)
- Mobile app (native iOS and Android for better voice input)
Long-term Vision:
- B2B partnerships: Companies use InterviewOS as a pre-screening tool
- Coach marketplace: Connect users with human coaches for follow-up sessions
- Interview platform: Host live interviews with feedback
- Career pivot assistance: Interview prep for career changers
- Salary negotiation coaching: Extend beyond interviews to offers
The Bigger Picture: The core mission is simple: Democratize interview coaching. Right now, wealthy people hire $200 per hour interview coaches, connected people have mentors who prepare them, and everyone else practices blind. InterviewOS levels the playing field. With Gemini 3's reasoning capabilities, we can provide world-class coaching to anyone, anywhere, for free.
Log in or sign up for Devpost to join the conversation.