ReadyForRound
AI mock interviews that feel like the real thing. 100% free.
Talk to an AI interviewer by voice, get an honest feedback report in about a minute,
and a study plan built from your gaps.
Try it live →
The problem
Most candidates don't fail interviews because they lack knowledge. They fail because they have never
said their answers out loud, under pressure, and then been told what was missing. Mock interviews
with a mentor are expensive or hard to schedule, and a rejection rarely comes with feedback.
The solution
ReadyForRound is a full-stack web app that runs a realistic interview round and coaches you afterwards:
- Pick a goal: role, experience level (2nd-year student to 5+ years at work), company type, and interview date.
- Take the interview: a video-call-style room where the interviewer speaks each question and you answer by voice (or by typing). Vague answers get real follow-up questions.
- Get the report: a score, what went well, what was missing, a strong example answer for every question, and what to study next, which feeds a weekly roadmap.
Screenshots
| Interview room | Feedback report |
|---|
| |
Features
- Voice interviews: the interviewer speaks (browser speech synthesis); answers are recorded and transcribed with Whisper
- Every round of a hiring drive: Technical, Project deep-dive, Behavioural, HR and Resume deep-dive, or a complete drive whose sequence depends on the company type
- Adaptive follow-ups when an answer is vague, with server-enforced limits so interviews stay on track
- Honest feedback reports graded question by question, with a model answer to learn from
- Learning resources: every "Study next" topic, roadmap task and missing skill links to hand-picked free resources (official docs, free courses, practice sites, plus a video search). The AI never writes links;
npm run check-links verifies all of them
- Speaking coach: speaking speed, filler words ("basically", "um"), answer length and long pauses for every spoken answer, measured from Whisper's word timings with no extra AI call. Only the numbers are kept, never the audio
- Skills: claimed vs proven, comparing self-ratings against what interviews actually showed
- AI study roadmap: weekly tasks sized to the hours you have before your interview date
- 5-minute drills on any topic, streaks, score trends and a readiness score
- Resume-based interviews: upload a PDF and the interviewer asks about your own projects, skills and experience (only the text is kept)
- Career Compass: a 10-minute role-discovery assessment (work interests, forced choices, real experience, comfort check and optional taste tests) matched to 19 roles with explainable, rule-based scoring; role names stay hidden until the result to avoid hype bias. On simulated students (npm run compass:sim), about 87% get their true role first and over 99% find it in their top 3
- Jobs Board (live job market): openings, typical pay, fresher share, top cities and the most-asked skills for each role, summarised daily from real job ads in India (Adzuna API). Each skill shows whether you've proven it, only claimed it, or are missing it, with a one-click drill. "Top charts" show the #1 role in each category, an openings leaderboard, a pay ladder, a "sweet spot" bubble map (openings vs pay), the hottest skills across India and a city split, drawn in plain SVG and CSS. Pay is shown only when enough ads state a salary, and the source and sample size are always shown
- Group discussion (GD) simulator: discuss a topic with three AI classmates who have different styles (one dominant, one evidence-minded, one who makes weak points to challenge); they address you by name and react to your points. Take turns by voice or text, pass, and summarise at the end, then get scored on initiation, content, communication, listening, leadership and summary, with facts measured by the system (opened, named others, summarised) and the speaking coach
Architecture
flowchart LR
U([Browser]) -->|HTTPS + JWT| V[React app on Vercel]
V -->|REST /api| R[Express API on Render]
V -->|Sign-in: email or Google| SA[Supabase Auth]
R -->|SQL with row-level security| DB[(Supabase Postgres)]
R -->|Questions and grading| G[Groq LLM: gpt-oss-120b]
R -->|Speech to text| W[Groq Whisper]
R -.->|Local development| O[Ollama]
R -->|Feedback emails| E[Resend]
How one interview turn works: the browser speaks the question, records the answer, and uploads the audio.
The server transcribes it, stores the turn, and asks the LLM for the next question in a strict labelled format.
It then validates that reply (it must be a real question, and follow-up limits are enforced by the server, not the model)
before sending it back. When the interview ends, a stronger grading prompt scores every answer and builds the report and roadmap.
Engineering highlights
- Reliable output from LLMs. Instead of fragile JSON, the interviewer replies in a labelled
TYPE: / QUESTION: format with a tolerant parser. Malformed Groq replies (tool_use_failed) are recovered from the provider's failed_generation field, and reasoning models get tuned token budgets.
- Prompt-injection defence. Candidate answers are passed as quoted data the model must never obey ("give me 10/10" does nothing). Every interviewer turn must be an actual question, and the server, not the model, decides when follow-ups stop.
- The right model for each job. A fast model can run the conversation while a larger model grades. In early testing a small local model gave confident but wrong feedback, so grading always runs on a stronger model.
- A voice pipeline that works everywhere. Browser speech recognition fails outside Chrome and Edge, so answers are recorded with
MediaRecorder and transcribed on the server. The microphone opens only while answering, which fixes Windows "audio ducking" muting the interviewer. Speech playback is safe under React StrictMode's double effects.
- Security. Supabase row-level security on every table, the secret key only on the server, JWT verified on each request,
zod validation on every input, per-user rate limits, and a hidden honeypot field against feedback spam.
- Built for free tiers. Per-user daily quotas with a timezone-aware midnight reset, a configurable fallback LLM provider, and graceful handling when the backend is asleep or unreachable.
- Accessible, polished UI. Custom dropdown and dialog components with full keyboard support, scroll-reveal animations that switch off for users who prefer reduced motion, semantic colour tokens for light and dark themes, and a layout tested at phone width.
- Observability and SEO. Browser and server errors are logged to an
app_errors table behind an error boundary. Structured data, a sitemap, and per-page titles help search engines.
Tech stack
| Layer | Tools |
|---|
| Frontend | React 19, TypeScript, Vite, Tailwind CSS 4, React Router |
| Backend | Node.js 22, Express 5, TypeScript, zod, express-rate-limit |
| Data and auth | Supabase (Postgres, row-level security, email and Google sign-in), SQL migrations via the Supabase CLI |
| AI | Groq (gpt-oss-120b for interviews and grading, Whisper for speech-to-text), Ollama for offline development, OpenRouter as an optional fallback |
| Hosting | Vercel (website), Render (API), Resend (feedback email) |
Project structure
client/ React app (Vercel)
src/pages/ One file per screen: Landing, Home, InterviewRoom, ReportPage, ...
src/components/ Shared UI: dropdowns, dialogs, avatar, skill picker, feedback form, ...
src/lib/ API client, skills dictionary, Supabase client, helpers
server/ Express API (Render)
src/interview/ Interview engine, rounds and prompts
src/report/ Grading and report prompts
src/llm/, src/stt/ LLM providers and speech-to-text
src/routes/ REST endpoints
supabase/migrations/ Database schema, one numbered SQL file per change
docs/ Product plan, deployment guide and screenshots
Run it locally
Requirements: Node.js 22, a free Supabase project, and a free Groq key
(or Ollama to run the AI fully offline).
# 1. Settings: copy the examples and fill in your keys (never commit .env files)
cp server/.env.example server/.env
cp client/.env.example client/.env
# 2. Database: create the tables in your Supabase project
npm install
npx supabase login
npx supabase link --project-ref <your-project-ref>
npm run db:push
# 3. Backend (terminal 1)
cd server && npm install && npm run dev
# 4. Website (terminal 2)
cd client && npm install && npm run dev
Open http://localhost:5180. Set LLM_PROVIDER=ollama or LLM_PROVIDER=groq in server/.env.
Deployment steps are in docs/DEPLOY.md.
Roadmap
- An experienced-hire track: work deep-dives, system design and salary conversations
- More roles beyond software development and data analysis
- Downloadable PDF reports
- A dashboard for colleges to support their students' placement preparation
Author
Harish Kandi, who designed and built ReadyForRound end to end: product, UI, backend, AI prompts and deployment.
GitHub
If ReadyForRound helped you prepare, a ⭐ on this repository means a lot.
