CallLens

Sales calls transcribed, scored by an LLM, searchable by meaning

Every sales call transcribed, scored, and searchable. CallLens ingests a call via signed webhook, transcribes it with diarization, scores each rep against a configurable scorecard with a structured-output LLM, builds embeddings for semantic search, and serves the whole thing through a Cube-powered analytics cabinet. Zero GPUs. Single Hetzner box.

Sector: AI · SaaS
Stack: Symfony 7.4 · PostgreSQL 17 · Next.js 15 · Cube

In numbers

GPUs Required: 0
Vector Dimensions: 1024
Pipeline Stages: 5
Monthly Infra Cost: ≈€60

01 — About the Project

Sales Coaching That Actually Reads the Calls

Most sales teams record their calls. Almost none get them scored consistently — managers don't have the hours, and scorecards don't enforce themselves. CallLens automates the read: every call ingested, every rep scored, every objection searchable.

We built it on the cheap-by-design principle: all heavy AI runs through external HTTP APIs (Deepgram for STT, OpenAI for scoring + embeddings), no GPU needed, a single Hetzner dedicated server is sufficient. Provider-independent interfaces let you swap or self-host later without touching business logic.

02 — The Challenge

Score sales calls reliably with an LLM that never hallucinates an evidence quote, persist 1024-dim utterance embeddings for semantic search at scale, give managers a Cube-powered analytics cabinet — and do it all on a single Hetzner server with one Docker stack and no GPU.

03 — The Solution

Symfony 7.4 + PHP 8.5 backed by PostgreSQL 17 + pgvector (HNSW), API Platform generating OpenAPI 3.1, a Messenger pipeline that walks each call through transcribe → score → embed states (idempotent, retryable), Cube as the semantic analytics layer, and a Next.js 15 cabinet on top. The scorer enforces structured JSON output and validates every evidence quote against the actual transcript.

04 — Key Features

Built for the Sales Manager

01
Webhook + Manual Ingest
HMAC-signed webhooks from any telephony/CRM (Twilio, Binotel, Ringostat, Bitrix24) plus manual upload via REST. Idempotent by tenant + external call ID.
02
Diarized Transcription
Deepgram for STT with per-channel labelling for dual-channel audio and provider diarization for mono. Word timings persisted on every utterance.
03
LLM Scoring with Evidence
Per-criterion scoring with strict JSON schema, an evidence quote per score, and the quote validated to appear in the transcript. No vibes scoring.
04
Semantic Search
1024-dim pgvector embeddings on every utterance, HNSW index, tenant-scoped ANN. Ask 'where did the rep handle objections' and get the actual moments.
05
Cube Analytics
Pre-aggregated dashboards on Postgres — avg score per agent, trend over time, top objections, score distribution. No second search engine for reports.
06
Configurable Audio Retention
keep, delete-after-processing, or delete-after-N-days, per tenant. Scheduler sweeps run nightly; transcripts and scores remain even after audio is deleted.

05 — Screenshots

See It in Action

CallLens — Mobile Version — Mobile Version

06 — Technology

Under the Hood

Single Docker stack, single Hetzner box, single Postgres. All heavy AI work goes through external APIs behind narrow interfaces — swap providers in a config line, or move to self-hosted GPU later when the cost curve crosses over (~215k audio-minutes/month).

Symfony 7.4: PHP 8.5 API + Workflow
PostgreSQL 17: + pgvector (HNSW)
Next.js 15: Landing + docs + cabinet
Cube: Analytics over Postgres

Need AI-Native Tooling for Your Team?

From sales coaching to support analytics, we ship AI products that get the heavy lifting right — strict structured outputs, evidence-grounded answers, and stacks you can actually operate.

Get in Touch →View All Projects

Next projectBeauty · Spain