All work
Voice AI
2024 – Present

CloserCoach

Voice-AI Sales Coaching Platform · $60K+ MRR

Tech stack

Node.jsExpressFirebaseVAPILangfuseOpenAI

Architecture highlights

  • VAPI real-time voice + Langfuse prompt observability
  • Firebase Realtime DB for live analytics
  • Retry-safe webhook workers for 50K+ daily events

The problem

Sales reps had no scalable way to practice and get objective feedback — human coaching didn't scale past a handful of calls per week.

What I built

I architected the backend and infrastructure from scratch (Node.js, Express, Firebase) for an AI sales-coaching product that runs realistic voice roleplays and returns instant, structured feedback.

Key features

  • Real-time voice roleplay via VAPI with < 1.5s feedback latency
  • AI analysis of 10,000+ daily sales conversations
  • Admin dashboard: analytics, user management, revenue tracking
  • Webhook pipeline handling 50,000+ callbacks/day with retry logic

The hardest challenge

Guaranteeing sub-1.5s feedback while processing tens of thousands of voice callbacks daily — solved with an idempotent, self-healing webhook pipeline and Langfuse-tuned prompts.

The full story

Problem

Human sales coaching doesn't scale. Reps needed unlimited, realistic practice with objective, instant feedback.

Research

I evaluated real-time voice providers and prompt-observability tooling, landing on VAPI for voice and Langfuse for prompt engineering and tracing.

Architecture

Node.js + Express APIs, Firebase Realtime Database for live analytics, and a retry-safe webhook worker layer to absorb 50,000+ VAPI callbacks per day.

Implementation

I built the entire backend and infrastructure from scratch: voice session handling, the feedback pipeline, an admin dashboard with revenue tracking, and idempotent webhook processing.

Result

The product sustains 500+ daily signups and $60K+ in monthly recurring revenue, with 99.9% webhook reliability and sub-1.5s feedback latency.

Lessons learned

At real-time voice scale, idempotency and retry logic aren't nice-to-haves — they're the product. Designing the webhook layer to self-heal was the single highest-leverage decision.

Outcome

  • $60K+ monthly recurring revenue enabled
  • 500+ daily signups sustained
  • 99.9% webhook reliability at 50K+ events/day

Want a result like this?

Let's talk about your project and the outcome you're after.

Start a project
Next case studyRAG Document Q&A