Gatekeeper AIReal-Time Voice Interview Coach
A real-time voice interview and sales coach: users practice out loud against an AI that speaks first, pushes back like a real interviewer, and scores every answer against the STAR+R framework - so they find their weak spots before the interview does.
Built with Next.js 14, React 19, TypeScript, Go, PostgreSQL, Google Gemini Live API, Deepgram, Hugging Face embeddings, Tailwind CSS, Vercel.
7 Scenarios
Job interview, salary negotiation, cold call, discovery, demo, closing, performance review
Voice-First
Real-time spoken role-play via Gemini Live and Deepgram - not a text chatbot
STAR+R Scoring
Semantic matching grades each answer on structure and surfaces the gaps
What it is
Most interview prep is silent and passive - rereading notes and rehearsing answers in your head. Gatekeeper AI turns it into live, spoken practice. You talk to an AI interviewer that opens the conversation, asks pointed follow-ups, and refuses to just nod along. Upload a resume and it tailors the questions to your background, scores each answer on the STAR+R framework (Situation, Task, Action, Result, Reflect) using semantic matching, and coaches you in real time with the Live Whisperer. The same engine powers seven scenarios spanning job interviews and sales conversations.
The problem
Interview and sales-call performance is a spoken skill, but almost all prep is written and passive. People rehearse in their heads, never hear themselves under pressure, and get no objective read on whether an answer actually landed - so they discover their weak answers live, in the room that counts.
What I built
Gatekeeper AI runs a real-time voice conversation. Google Gemini Live API drives a professional interviewer persona that speaks first and pushes back, while Deepgram streams speech-to-text so answers are transcribed and scored as you talk. Each response is graded against the STAR+R framework with client-side semantic embeddings, a Live Whisperer surfaces coaching mid-call, and a drill mode lets users hammer their weakest questions. A Go and PostgreSQL backend stores sessions and scored assessments; a Next.js 14 and React 19 frontend delivers the low-latency voice UI. Seven scenarios cover job interviews, salary negotiation, cold calls, discovery, demos, closing, and performance reviews.
What happened
Live in production with the full practice loop - scenario selection, resume-tailored prep, real-time voice or text sessions, STAR+R answer scoring, and post-session assessments - deployed on Vercel with automated GitHub deploys.
What it changed
- Lets job seekers rehearse out loud against realistic interviewer pushback before the real thing
- Covers interview prep and sales-conversation training in one tool
- Turns passive prep into measurable practice with scored, replayable sessions
Who it fits
- Job seekers practicing behavioral and technical interviews out loud
- Sales reps rehearsing discovery, demo, cold-call, and closing conversations
- Anyone prepping for a salary negotiation or performance review
- Real-time bidirectional voice via Gemini Live API with handle-based session resumption
- Deepgram Nova-2 streaming speech-to-text with diarization
- STAR+R answer scoring using client-side Hugging Face semantic embeddings
- Live Whisperer: real-time coaching suggestions generated mid-call
- Go (net/http) and PostgreSQL backend behind a Next.js 14 and React 19 frontend
Want something like this built?
Bring the idea or the app. Step 1 is a free 15-minute fit call. Step 2 is a paid discovery session, $500, that ends in a written plan. Step 3 is a fixed quote for the build, or you take the plan and build it yourself.