A system that listens to every call and stands behind its numbers
Revizor connects to the working CRM (here, the on-premise edition of Bitrix24) and pulls in everything: deals, leads, call recordings, messenger chats. The neural network transcribes conversations by role (“rep / customer”), breaks every deal down into 53 structured facts (need, city, objections, reason for loss, promises) and turns them not into “pretty charts” but into answers: where we're losing, who's losing, what to do today.
The system's main rule: every number is clickable. A metric drills down to a list of deals, a deal to the conversation transcript, an AI conclusion to the customer's own words. People trust a report they can verify in two clicks, and they make decisions based on it.
What management sees every morning
One screen instead of a stand-up
Today's priorities from AI
- For every active deal, what to do right now: call back, push to close, send a quote
- Sorted by importance: the head of sales hands out tasks in 5 minutes instead of spending an hour on deal cards
- A customer summary right in the row, without opening the CRM
Said it, did it
Every conversation under control
- AI checks every conversation against a 17-point sales checklist, with quotes
- Promises to customers are tracked: if a rep says “I'll send the estimate today” and doesn't, you see it right away
- The score is built on the rep's own words, not on the manager's impression
Advertising without the spin
Traffic quality based on facts
- Every inbound request is scored on a 100-point scale: volume, location, whether there's a contact
- The rep's work doesn't affect the lead score, so the ad agency has nowhere to hide
- A drop in quality shows up the day it happens, not in an end-of-month report
Geography and product
Demand, as customers voice it
- A demand map by city with the dealer network: where to put ad spend, where to open a warehouse
- Which collections and colors customers name out loud: the structure of demand for production and purchasing
- Complaints from conversations, sorted by defect category, without a single survey
How it works under the hood
CRM → data mirrors
46 tables from a MySQL replica of Bitrix24: nightly reload, change detection, self-healing after failures
Ears: WhisperX
Transcription of calls and voice messages with speaker diarization by role, on a GPU, with no per-minute fees to external services
Brain: LLM pipeline
Claude / DeepSeek with automatic failover, queues and caching: 53 fact fields with quotes, for pennies per deal
Data marts and web app
39 PostgreSQL data marts, 33 pages in Next.js: reports, a management cockpit, a pipeline monitor
Why even the reps trust the numbers
A key design decision: AI returns only facts with quotes, and the scores come from an open formula with weights that management adjusts in the settings without a developer. No “AI opinion”: behind every score there's a line from the conversation and simple arithmetic. Weight changes apply instantly, without re-running the analysis.
