Overview

nextneural_intel analyses call recordings after the fact. You upload audio; it transcribes, extracts structured fields, scores the conversation, and rolls the results into dashboards and reports.

Unlike nextneural_converse, it never places or answers a call. Its input is a recording of a call that already happened, wherever it came from.

How the pieces fit

Project ──► Upload ──► Call ──► Analysis ──► Dashboard / Reports
   │                     │          │
   │                     │          └─ extraction fields, scores, signals
   │                     └─ Customer ──► Events, phones, segments
   └─ call_context: what this project's calls are about

1. Projects partition the workspace; one is active at a time
2. Uploads bring recordings in, singly, in bulk, or by CSV
3. Calls are the analysed recordings
4. Customers accumulate across calls, with their own history
5. Segments group calls and customers for reporting
6. Reports are generated documents you can send on

Base paths

Surface	Base	Credential
Console API	/api	Session token
Developer API	/api/v1/sales	API key (public API)
Simulation judging	/api/nn-intel/simulations	API key, used by nextneural_coach
Auth	/api/auth	—

Interactive API docs

nextneural_intel generates its own OpenAPI document from its live route table:

Endpoint	What
GET /openapi.json	The generated spec
GET /docs	Swagger UI

Generated from the real routes, so it cannot drift from what the binary actually serves. It describes paths and methods; the request and response detail is here.

The analysis pipeline

An uploaded recording moves through these statuses:

Status	Meaning
pending	Queued, not yet picked up
awaiting_recording	The row exists but the audio has not arrived
transcribing	Speech recognition in progress
completed	Analysed — transcript, extraction, and scores are available
failed	Something in the pipeline failed; the reason is on the call

Work is queued through River (a Postgres-backed job queue), so a restart does not lose queued analyses — they resume when the binary comes back.

A failed call can be retried without re-uploading the audio.

Extraction

What gets pulled out of a transcript is configurable rather than fixed. Extraction fields define the questions asked of every call — outcome, disposition, next action, whatever your business needs. Editing them is a superadmin action, since they apply across the deployment. See Settings.

Built-in call outcomes: resolved, follow_up_scheduled, visit_scheduled, escalated, interested, no_interest, voicemail, failed.

Requirements

For	Requires
Building the binary	A C toolchain — the VAD binding needs cgo
Running it	ffmpeg on PATH (or FFMPEG_BINARY)
Analysis	A configured model route — see Settings
Reports by email	RESEND_API_KEY / RESEND_SENDER_EMAIL

Identifiers

Unlike the other three binaries, nextneural_intel's console routes address resources by integer id, not UUID:

GET /api/calls/1042
PATCH /api/customers/87

UUIDs still exist on the records and are what the public API returns.