Call Analytics Platform — Automated Transcription & AI Quality Scoring for a Sales Call Center (Go)

Overview

The client operates a sales/support call center and was drowning in raw audio. Recordings landed on an FTP server, but supervisors had no way to know — without listening to hundreds of MP3s per day — which calls were good, which violated the script, or which customers were unhappy.

They needed a single internal web application that would automatically ingest every call from their telephony provider, transcribe it, evaluate it against a configurable list of quality questions, and present the results in a fast, filterable dashboard their managers could actually work with.

My Contribution

I designed and built the entire system end-to-end in Go.

Backend & Pipeline

A webhook endpoint (POST /call) ingests call metadata pushed by the telephony provider and enqueues background jobs. Five independent background workers, each backed by a database-driven queue with pending → done / failed status tracking, attempt counters, retry logic, and per-job error capture:

• Duration worker — connects to the client’s FTP server, streams the MP3, and extracts true audio duration.
• Transcription worker — sends audio to AssemblyAI and stores the transcript with a “too-short” guard.
• AI evaluation worker — composes a prompt from the transcript plus a dynamic list of quality questions managed from the UI, calls OpenAI, validates the structured JSON response, and stores it with a hash for change detection (so re-evaluation only happens when inputs actually change).
• Webhook log cleanup worker — applies configurable retention.
• Error log cleanup worker — same, for error history.

Web App

• HTTP layer built on Echo v4 with a clean handler → service → repository split.
• Server-side rendered UI using Templ (compiled, type-safe Go templates) — no SPA overhead, fast first paint.
• DataTables with full server-side processing (pagination, filtering, sorting done in PostgreSQL, not the browser) for the calls table and webhook log table.
• Per-user column visibility persisted in the database, so each manager sees their own preferred view.
• Role-based access control (admin / regular user), Redis-backed sessions via scs, secure password change flow.
• In-app settings UI for AssemblyAI keys, OpenAI model/parameters, AI questions, and log retention — no redeploys needed to tune the pipeline.
• Built-in MP3 audio playback proxied from FTP, so managers can listen to a call right next to its transcript and AI score.

Infrastructure

• PostgreSQL (sqlx) with 25+ versioned migrations applied automatically on startup via golang-migrate.
• Redis for sessions.
• Docker Compose for local development; single binary for production.
• Timezone-aware (Europe/Kyiv), structured logging via logrus.

In-App Configuration

Every part of the pipeline is tunable from the UI — no redeploys. Supervisors manage AssemblyAI keys, OpenAI model parameters, minimum audio duration, transcription routes, and log retention directly in the settings screen.

The list of AI quality questions is itself a CRUD-managed entity. Adding a new criterion (“did the operator greet?”, “was a conflict raised?”, “was a TG-bot proposed?”) takes seconds: type it into the questions UI, and the next batch of calls is evaluated against it.

Technical Stack

Outcome

The client replaced hours of manual call listening with an automated pipeline that processes every recorded call within minutes of it landing on FTP.

• Supervisors now open a single dashboard, filter by date, agent, or AI verdict, and drill straight into a transcript with per-question scoring.
• Adding a new quality criterion takes seconds — they type the question into the settings UI and the next batch of calls is evaluated against it.
• The platform has been running reliably in production and is the team’s primary tool for call quality review.