AI Supervisor Module

AI analysis of MikoPBX call transcripts

The AI Supervisor module analyzes completed MikoPBX call transcripts and helps supervisors monitor communication quality. It produces short and detailed summaries, identifies the contact reason and call outcome, highlights next actions, topics, risks, and supporting evidence, and evaluates sentiment, risk, and service quality.

The MikoPBX module does not run a language model itself. It imports transcripts, manages the queue, and stores access keys and results. A separate AI Supervisor Worker application performs the local processing on a Mac. By default, the application uses Ollama and a local model, so transcripts and analysis results remain within your infrastructure.

Module home page

How analysis works

The solution consists of three independent components:

1. The Local Speech To Text module creates a transcript and publishes the transcript.completed event.

2. The AI Supervisor module imports the transcript, creates a job, and saves the analysis result.

3. AI Supervisor Worker on the Mac receives the job, prepares the local model, performs the analysis, and sends the result to MikoPBX.

A typical processing cycle is as follows:

1. After a call, the speech recognition module creates a completed transcript.

2. AI Supervisor imports the event and saves a reference to the transcript.

3. A call_summary job is created for an eligible call.

4. The worker registers and acquires a lease through POST /job-leases.

5. For a long conversation, the worker analyzes it in chunks and saves partial results.

6. The final JSON is validated in the application and then validated again in the module.

7. The result appears on the Overview and Calls tabs.

Requirements and compatibility

• MikoPBX 2025.1.1 or later.

• The Local Speech To Text module installed and enabled.

• Network access from the Mac to the MikoPBX web interface.

• Internet access for the initial Ollama installation and local model downloads.

Installing the module

1. Open the MikoPBX web interface.

2. Go to Modules → Module marketplace.

Module marketplace

1. Find AI Supervisor module and install it.

2. Open the Installed modules tab and enable the module.

Enabling the AI Supervisor module

1. Click the settings button to the right of the module version.

Module settings button

Module navigation

The current version has three main tabs:

Tab	Purpose
Overview	Call metrics, AI analysis coverage, the attention queue, and aggregate analytics.
Calls	Imported calls, filters, the call card, transcript, analysis, and case workflow.
Settings	Workers, import, analysis components, attention rules, catalogs, the queue, and diagnostics.

Overview tab

The Overview tab shows call supervision status for the selected period.

Overview tab

Key metrics:

Metric	Meaning
Calls with transcripts	Number of imported calls during the selected period.
AI analyzed	Number and proportion of calls with a saved analysis result.
Avg. call duration	Average duration of calls during the selected period.
Needs your attention	Calls not yet reviewed by an operator that match risk, quality, sentiment, or AI problem-flag rules.

Below the metrics are charts for calls over time, employees, call directions, communication tone, and calls that need review. Selecting a metric or analytics block opens the corresponding list on the Calls tab.

Calls tab

The Calls tab is the supervisor's main workspace.

Calls tab

You can search, sort, and filter by period, direction, employee, sentiment, quality, risk, completion status, workflow status, due date, and AI analysis state. The AI analysis state can separately show calls that are in progress, complete or incomplete, waiting for speech recognition, or not yet started.

Workflow statuses:

Status	Purpose
New	The case has not yet been taken into work.
In progress	A supervisor is reviewing the case or waiting for a follow-up action.
Processed	The review is complete.

You can select multiple rows and use bulk actions to take cases into work, mark them as processed, or return them to New.

Call card

The call card combines call details, built-in recording playback, the transcript, analysis results, and the case workflow. It includes:

• short and detailed summaries;

• contact reason, outcome, and next actions;

• topics, risks, and supporting evidence;

• risk, quality, and sentiment scores;

• key moments with time navigation;

• voice metrics and emotions when the corresponding stages are enabled;

• the state of every AI analysis stage;

• assignee, due date, escalation, and notes;

• case workflow history.

Call card

Processing queue

The queue is located under Settings → System → Processing. It shows active, failed, completed, and skipped jobs. You can refresh the data, retry an individual failed job, retry all failures, return expired leases to the queue, or clear the unfinished queue.

The waiting_for_stt state means that a job needs a transcript, but neither an available snapshot nor a response from the speech recognition module is available. This waiting state is not an error: the job is not leased to a worker and does not consume attempts.

Settings tab

Settings are divided into six work areas. Changes are saved automatically; there is no separate Save settings button.

Workers

Create and delete access keys for AI Supervisor Worker here. The complete value of a new key is displayed only once. Copy it before closing the window.

Worker access in module settings

Call flow

This section configures:

• automatic AI analysis;

• processing of internal calls;

• contact names from PhoneBook;

• AI result language;

• automatic import and manual import runs;

• transcript and result retention: 3 months, 6 months, 1 year, 2 years, or forever.

Call flow section

AI analysis

This section defines the analysis pipeline and the models to be used:

LLM stands for Large Language Model. It is a machine-learning model that can work with language: read text, understand its meaning in context, and generate coherent responses.

Component	Purpose
Summary (LLM model)	Structured call review, topics, risks, and quality.
Voice metrics (Technical analysis)	Tempo, pauses, interruptions, silence, and speaker balance.
Emotion Analysis (LLM model)	Text-based emotion detection for selected fragments.
Acoustic analysis (Acoustic model)	Asynchronous detection of acoustic features in the recording.

Current LLM model profiles:

Profile	Model	Approximate memory	Purpose
Qwen3 8B Instruct Q4	qwen3:8b	6–9 GB	Recommended balanced profile.
Qwen3.5 4B Q4	qwen3.5:4b	4–6 GB	Fast, compact profile.
Qwen3.5 9B Q4	qwen3.5:9b	8–12 GB	Detailed analysis of long and complex calls.

Expert parameters include the time the model remains loaded in memory, request timeout, and number of attempts. The selected profile determines the context and chunk sizes.

Module AI analysis parameters

Attention rules

Set risk and quality thresholds and rules for negative sentiment and AI problem flags here. Changes affect the review queue and filters; saved model responses are not recalculated.

Attention rules and sensitivity settings

Search rules

This area contains two tabs:

• Important situations — system and custom call risk types, their importance for analysis, and instructions that tell the model how to recognize each situation.

Search rules → Important situations

• Call topics — company-specific topics and aliases that normalize different AI-generated descriptions under the same topic.

Search rules → Call topics

System

The System area contains:

• Status — readiness of speech recognition, worker access, import, and the queue;

• Processing — jobs, retries, expired calls, and unfinished call queue cleanup;

• Diagnostics — logs and a technical debugging summary.

Former diagnostics screen

REST API

Base path:

/pbxcore/api/v3/module-ai-supervisor

Method	Address	Purpose
GET, PATCH	/settings	Read and update settings.
GET	/dashboard	Overview data.
GET, PATCH	/calls, /calls/{id}	Call list, call card, and workflow updates.
GET, POST	/call-notes	Read and add notes.
GET	/call-workflow-events	Workflow history.
GET, POST, PATCH, DELETE	/jobs, /jobs/{id}	Queue and job actions.
GET, POST, DELETE	/worker-api-keys, /worker-api-keys/{id}	Worker access keys.
POST, GET	/imports, /imports/{id}	Start an import and read its status.
GET	/logs	Module log.

Primary worker resources:

• GET /worker-api-contract;

• POST /workers;

• POST /job-leases;

• PATCH and DELETE /job-leases/{id};

• GET /job-recordings/{id};

• PUT /job-results/{id};

• PUT /job-failures/{id};

• PUT /job-partials/{id};

• PATCH /job-voice-analytics/{id}.

Troubleshooting

The Calls tab is empty

• Check that the Local Speech To Text module is installed and enabled.

• Make sure it contains completed transcripts.

• Open Settings → System → Status.

• Check automatic import under Settings → Call flow and run an import manually if necessary.

Jobs are not being processed

• Open Settings → System → Processing.

• Check for errors and the waiting_for_stt state.

• Make sure automatic analysis is enabled under Call flow.

• Enable the corresponding setting if internal calls must be processed.

The worker cannot connect

• Check compatibility between ModuleAISupervisor 1.73 and AI Supervisor Worker 1.7 build 34.

• Make sure you are using an AI Supervisor key, not a Local STT Worker token.

• Check the PBX address, TLS certificate, and CA file.

• Open Diagnostics in the application and run the connection test.

The model cannot be prepared

• Open Models in AI Supervisor Worker and click Refresh.

• Check Ollama status, the model selected by the PBX, and free disk space.

• Select Qwen3.5 4B Q4 to use less memory.

• See Activity and Diagnostics for download details and errors.