Your operation's analyst, your agent's coach
CALLAII - Callmenta Analysis & Learning · Linguistic AI Insight. More than an analytics tool: it's your team's coach, trainer and advisor. Driven by an operational LLM, it talks to every agent, generates feedback tied to their own performance, and ships reports straight to the quality team.
- Talks one-to-one with every agent and delivers personalised feedback.
- Generates daily micro-coaching off the back of each call.
- Surfaces live risk and opportunity reports to supervisors.
The figures are typical ranges taken from pilot studies and reference customer data. They vary by segment and operation and do not guarantee your own result.
CALLAII Pipeline
Approved evaluations become training data
Every approved soft-skill evaluation exports in one click to ChatML or Alpaca. A ready package for fine-tuning your own LLM on your own operational data.
- Call transcript + rubric score collapsed into a single line
- Human and AI agent samples labelled separately
- Instant export to JSONL, ChatML or Alpaca
- Train your own model on your own dataset
A bot does not answer the phone
CALLAII does not put a bot in front of your customer on the phone. A person always handles the conversation and the artificial intelligence stands behind the agent. On written channels a person answers during working hours too.
What CALLAII does
- Produces the transcript as soon as the conversation ends
- Tags the topic, the outcome and the sentiment
- Calculates the quality score, so nobody listens to a hundred calls
- Shows prompts on screen while the agent is talking
- Writes the coaching task and the end-of-day summary
- Prepares the reports and sends them on time
What CALLAII never does
- It does not take your customer's call on the phone
- It does not answer in place of the agent during working hours
- It does not pretend to be human, it always says it is an assistant
- It never removes the option of reaching a person
- It does not send a single email without your approval
The after-hours assistant add-on covers your written channels outside working hours. It answers from your own knowledge base, introduces itself as an assistant in its first sentence and hands over in the morning what it could not solve. It does not work on the phone. 4,900 TRY a month, optional.
What does CALLAII actually do?
Every capability was built in answer to a real problem on the floor.
Speech → Meaning
Turkish-native ASR (Whisper-TR fine-tune). Accent, jargon, phone-line audio quality - trained on real call data.
Intent & Context
Not what the customer said - what they meant. The NLU layer pins down call intent: sale, complaint, cancellation, info request.
Emotion & Risk
Turkish-native emotion model. Maps anger, satisfaction and risk from tone and word choice.
QA Score
Script alignment, soft skill, KVKK, sales cues. Explainable scoring on a 100-point scale with evidence-backed feedback.
Personal Coaching
Daily micro-coaching for every agent, targeted at their own development area. Sample lines, alternative responses, focused assignments.
Live Action
High-risk calls pushed to a manager in real time. Escalations, returns and churn - caught before they happen.
RAG Knowledge Base
Searches your SOPs, product docs and customer records. The agent and the AI speak from the same source.
Multichannel
Voice, email, web chat - one pipeline. Context and customer profile follow the conversation across channels.
Analytics & Executive Summary
Daily operations digest, team performance dashboard, trend analysis. Whatever a manager asks, CALLAII answers.
Watch the operation live in one screen
Call coverage, soft-skill score, risk distribution, automated actions - the exact dashboard a manager opens at 08:30.
- Low · %64
- Medium · %22
- High · %10
- Critical · %4
- Empathy 87.4
- Ownership 82.1
- KVKK check 95.2
- Persuasion 78.6
- 09:14 Critical risk escalation 3
- 10:02 Coaching plan generated 12
- 12:48 Callback triggered 38
- 14:30 KVKK closure 6
The screen above is a representative example, not real customer data.
Not only calls, every chat too gets analysed
Alongside voice calls, CALLAII reads WhatsApp, Instagram, Facebook and website live chat at the same depth. For every conversation it produces the sentiment, the intent, the satisfaction score, the summary and the suggested action. It brings risky chats forward on its own.
Voice and text come together in one panel. Channel breakdown, satisfaction and agent performance are reported side by side.
CALLAII stands for
The abbreviation wasn't chosen for decoration. Each of the six letters represents a function the system actually performs.
Callmenta
The model is ours. Training, weights, and the server it runs on are all in Callmenta's hands; it doesn't connect to any external service.
Analysis
It analyzes the conversation: topic, request, objection, risk, and outcome. Same framework whether it's a voice recording or a written channel.
Learning
Every corrected decision becomes training data for the next version. The system narrows its own errors as it's used.
Linguistic
Trained on Turkish's own structure. Suffixes, inverted sentences, and field jargon are understood directly, not through translation.
AI
Not a rule engine. It evaluates situations that can't be caught by templates, and delivers its decision with reasoning.
Insight
The output isn't a scorecard. It tells the agent what to do and the manager where they're losing.
CALLAII was built by Callmenta
CALLAII is not an open-source assistant kit. It was distilled from very large foundation models and fine-tuned on the Turkish contact center floor using our academic library, industry papers, real call scenarios and audio recordings we produced ourselves.
- 01Academic library
Current papers, theses and an industry library on NLP, speech recognition and customer experience - the model's reference pool.
- 02Founders' own content
Thousands of training assets, articles and industry scenarios produced by our founding partners on contact center management, soft-skill coaching and quality frameworks.
- 03Turkish ASR voice fine-tune
CALLAII ASR was trained to native-level Turkish using real calls from dozens of industries and scripted audio recorded by our founding team.
- 04LLM distillation
From very large base models down to compact models aligned to contact center work - runs at 12 ms latency, even on edge hardware.
- 05Continuous calibration
Re-tuned per customer engagement - it learns your jargon, product names and SLA framework. Never one-size-fits-all.
Wrong decision does not stay in a blind spot
No evaluation comes from a single source. The rule layer and CALLAII make decisions independently; if they can't agree, the result isn't considered final and goes to a human. The human's decision overrides automation, and that decision feeds into the next training cycle.
-
Rule layer
Defined phrases are searched for verbatim in the transcript. If found, they are marked with the evidence and timestamp.
-
CALLAII opinion
The same question is asked to the model separately. The model must justify its decision with a verbatim quote from the transcript.
-
Contradiction gate
If two layers disagree, the result is marked as uncertain. An uncertain result does not trigger an alarm and does not assign a score to anyone.
-
Human review
Uncertain and negative decisions are queued for review. The agent's right to appeal is preserved in the same row.
-
Feedback loop
A decision corrected by a human becomes labeled data. The next version is trained to avoid repeating that error.
There is no negative decision without evidence. If the model quotes a sentence that does not appear in the transcript, that decision is considered invalid.
Humans on stage, in the kitchen CALLAII
A human always speaks with the customer. Behind the scenes every call is scored within hours, every agent gets personal coaching, and high-risk calls are pushed to a manager while they're still live.