The problem
Contact-center QA teams manually review a tiny sample of calls, so most interactions are never scored, coaching arrives days late, and compliance failures surface only after the fact.
Enterprises want AI to evaluate every interaction at volume, but they won't trust scores they can't audit, calibrate against human reviewers, or trace back to specific moments in a transcript.
The approach
Two complementary capabilities: real-time speech analytics that classifies a live call as it happens and pushes coaching to the agent's screen, and an automated QA layer where AI agents governed by Standard Operating Procedures score interactions against configurable scorecards and route the low scorers to human analysts.
A two-tier model strategy keeps it both fast and thorough — a fast model drives sub-two-second live coaching while a heavier model produces deep post-call analysis — and every AI score is calibrated against human reviewers and linked to transcript evidence so it stays auditable.
Automated workflows turn scores into action: escalations, coaching assignments, and audits fire off routing rules, with the AI handling volume and humans owning judgment and calibration.
What I built
- Real-time call classifier with an 8-stage sales-call funnel and 84+ diagnostic flags across six categories (compliance, sales process, communication, product knowledge, objection management, customer experience), severity-weighted into a 0-100 score with auto-fail flags for regulatory breaches
- Live coaching delivered to the browser over server-sent events during an in-progress call, backed by AI-driven telephony and a two-tier Gemini setup (Flash for real-time, Pro for post-call reports)
- SOP-governed automated QA engine on a unified Work-Unit data model: AI agents pick up evaluations, score against scorecards, and route sub-threshold results into a human analyst queue with downstream escalation/coaching/audit workflows
- Analyst tooling for AI-vs-human calibration, criterion-level rationale with transcript moments, human override with recorded reasoning, and dispute handling
- A rich analytics surface (16 custom D3 visualizations) plus an AI orchestrator chat that answers questions against live simulation context and flags ripple effects
- A no-login demo experience: a persona picker across six contact-center roles and an AI bootstrap wizard that generates a full quality program (scorecards, sample interactions, agents, workflows) for a chosen industry in seconds
- TypeScript
- Next.js 16
- React 19
- tRPC
- Prisma
- PostgreSQL
- Google Gemini (Flash + Pro)
- Vapi telephony
- Server-Sent Events
- D3.js
- Playwright
- Railway / Vercel
Result
A quality-management stack that scores interactions at volume instead of by sample — real-time on live calls and automatically after the fact — while keeping humans in the loop through calibration, override, and full scorecard-to-coaching traceability.
- Public interactive demo of the quality-management platform (persona picker, evaluation workflow, AI bootstrap wizard)
- 84+ severity-tiered diagnostic flags across 6 categories with a 0-100 composite score
- 16 custom D3 visualizations and an AI orchestrator over a simulated full contact-center day