Every call audited,
not one in fifty.
A QA team listening by hand gets through about two calls in a hundred. Nirikkhon AI transcribes all of them in Bangla and English, scores each one against your own SOP, and shows the line in the transcript that earned the score.
A score you can check against the recording
Each answer carries the words that produced it and the second they were said. Click a timestamp in the console and the recording seeks there, so a disputed score takes about fifteen seconds to settle.
আসসালামু আলাইকুম, কাস্টমার কেয়ার থেকে রুমানা বলছি। আপনাকে কীভাবে সাহায্য করতে পারি?
আমার ইন্টারনেট প্যাকটা কাল রাত থেকে চলছে না।
দুঃখিত। অ্যাকাউন্টটা দেখার আগে নিরাপত্তার জন্য আপনার জন্মতারিখটা বলবেন?
বারোই মার্চ, নিরানব্বই।
ধন্যবাদ। প্যাকটি চালু আছে, তবে ডেটা শেষ হয়ে গেছে।
Diarised and timestamped on the way in. Talk time, hold time and the agent's name are read off the same pass.
- Yes Opened with the approved greeting and gave their name 10
- Yes Verified the customer's identity before opening the account Evidence at 0:19 20
- Yes Explained the fault in plain language, without jargon 15
- No Offered the matching recharge before closing the call 15
Scored against Broadband — fault handling, the SOP matched to this agent's skill. A question marked fatal takes the whole call to zero.
Five stages, each one retried on its own
A recording lands, and the queue takes it from there. If a stage fails it retries by itself without redoing the stages before it, and you can watch every job move in the console.
-
1
Transcription
Who spoke, when, and for how long — with hold and talk time split out.
-
2
Classification
Why the customer called, how it ended, and whether a sale or a cancellation was in play.
-
3
Analytics
Sentiment turn by turn, plus the promises, dates and figures the agent committed to.
-
4
Audit
The weighted scorecard, answered against your SOP with the evidence attached.
-
5
Summary
A few factual lines a supervisor can read instead of the whole transcript.
Afterwards it looks for repeat contacts from the same customer and runs your alert rules over the result.
What your team works in
Getting calls in
- One recording at a time, or a whole folder with a metadata file beside each file
- A watched directory on the server: drop files in, they queue themselves, originals move to a done folder
- Chat and email threads, pasted or uploaded as text
- Re-scanning is safe — a recording already ingested is recognised by its contents and skipped
- A processing queue you can watch, with per-stage retries and stale jobs reclaimed automatically
Quality
- An SOP library of scripts, probing checkpoints, mandatory information and fatal notes, matched to each call by skill
- Weighted yes/no forms, where "not applicable" earns full weight and a fatal answer zeroes the call
- An audit workspace with the recording and transcript in step, and click-to-seek timestamps
- Your reviewers can re-score inline, and the call is rescored on save
- Sampling rules — so many calls per agent per period, spread across shifts, with the gaps named
- Calibration: how far the automated score sits from your reviewers', question by question
What you learn from it
- A dashboard of volume, verdict mix, sentiment, SOP usage and the leaderboard
- Agent profiles: score trend, pass rate per parameter, skill mix and a coaching brief
- Sentiment by agent and by call, with a negative-sentiment trend
- 23 reports across operations, agents and quality, every one exportable as CSV
- Training needs worked out from audit data, assigned, then measured again afterwards
- Alerts for regulatory wording, abuse, escalation language, long calls and fatal outcomes — thresholds editable by your own staff
Running it
- Users placed in a centre, a team and a line of business
- Roles with a full permission matrix, and a data scope — whole organisation, own centre, own team, own records — that every query obeys
- Your own reference data: centres, teams, lines of business, skills, shifts and brands
- An audit trail of every administrative action and every sign-in
- Password policy, lockout and session timeout you set yourself
Built to run inside your network
- Nothing calls out for assets
- Fonts, scripts and stylesheets are served by the application itself. No CDN to allow through the firewall.
- Bangla first
- Every font stack ends in a Bengali face, so transcripts render the same on a locked-down desktop as they do here.
- Your name on it
- Logo, colours, typography and the words the product uses for its own nouns — all changed in the console, with no rebuild.
- More than one operator
- Each tenant keeps its own users, SOPs, scorecard, prompts and data, and answers on its own hostname.
- Your choice of model
- Gemini, OpenAI or OpenRouter, set per tenant and per stage. A built-in offline provider runs the whole thing with no network at all.
- Every prompt is yours to edit
- Transcription through coaching, editable in settings with the previous version kept. Changes apply to the next call, never to results already recorded.
Have a look at a week of real calls
The demo tenant is loaded with processed interactions, audits, alerts and re-scored calls, so every screen has something in it.