Skip to content

Latest commit

 

History

History
170 lines (111 loc) · 8.17 KB

File metadata and controls

170 lines (111 loc) · 8.17 KB

Syntora Technology Stack

Syntora combines multiple AI and data technologies into a unified pipeline that ingests sales communications, understands them in context, and produces actionable recommendations. This page explains each technology layer and why it was chosen.

Architecture Overview

The data flow through Syntora follows this path:

Calls / Messages / CRM Data
        ↓
  Omnichannel Ingestion
        ↓
  Speech-to-Text (for calls)
        ↓
  Deal Context Assembly (RAG)
        ↓
  Dual LLM Analysis
        ↓
  Finding Aggregation & Scoring
        ↓
  Weekly Action Plan Generation

Each layer is purpose-built for sales communication analysis. Here's what each does and why.

Speech-to-Text Pipeline

Technologies: Nexara, OpenAI Whisper

Every phone call needs to be transcribed before analysis. Sales calls present unique challenges compared to general speech recognition:

  • Noisy phone audio — mobile connections, speakerphone, background office noise
  • Domain vocabulary — product names, pricing terms, industry jargon specific to each client
  • Two-speaker diarization — reliably separating the sales rep's speech from the customer's
  • Russian language optimization — most general STT models are tuned for English first

We use a dual-engine approach: Nexara for Russian-optimized recognition and Whisper for cross-validation on ambiguous segments. This catches transcription errors that a single engine would miss, especially on names, numbers, and technical terms that matter most in sales conversations.

Dual LLM Analysis

Technologies: OpenAI GPT, Google Gemini

Syntora uses two large language models in parallel rather than relying on a single one. This is a deliberate architectural choice for three reasons:

Why Two Models?

  1. Cross-validation — When both models agree on a finding (e.g., "the rep failed to handle a price objection"), confidence is high. When they disagree, the system flags the interaction for deeper review rather than making a potentially wrong call.

  2. Reduced hallucination — LLMs can confidently state things that didn't happen. Running two independent models and requiring consensus dramatically reduces false positives in scoring.

  3. Complementary strengths — Different models excel at different aspects of conversation analysis. One may better detect emotional tone, while the other catches logical gaps in objection handling.

What Each Analysis Evaluates

Dimension What the AI Looks For
Script adherence Did the rep follow the recommended conversation structure for this deal stage?
Objection handling Were objections acknowledged, explored, and addressed — or dismissed/ignored?
Upsell and cross-sell Did the rep offer relevant additional products? Used the right framing?
Next step commitment Did the call end with a concrete next action, or vague "I'll call you back"?
Emotional tone Was the rep confident, engaged, and empathetic — or rushed, bored, or defensive?
Compliance Were mandatory phrases spoken? Were prohibited topics avoided?
Deal risk signals Does this interaction suggest the deal is at risk of falling through?

Each dimension produces a score and, when issues are found, a specific recommendation tied to that rep and that deal.

RAG with Qdrant

Technology: Qdrant vector database

RAG (Retrieval-Augmented Generation) is what makes Syntora's analysis context-aware rather than generic. Here's why this matters:

The Problem with Context-Free Analysis

A generic AI analyzing a single call in isolation might flag "the rep didn't introduce themselves" — but if this is the 4th call with an existing client, an introduction would be awkward. Context matters.

How RAG Solves It

Before analyzing any interaction, the system retrieves relevant context from the vector database:

  • Deal history — all previous calls and messages in this deal
  • CRM data — deal stage, amount, products discussed, previous objections
  • Successful patterns — how top performers in this client's industry handle similar situations
  • Company-specific scripts — the recommended approach for this deal stage, customized for this business

This means the AI doesn't just evaluate "was this a good call?" — it evaluates "was this the right approach for this specific customer, at this deal stage, given what happened in previous conversations?"

Why Qdrant?

Qdrant is a high-performance vector database that enables semantic search — finding similar situations by meaning, not just keywords. When the system needs to find "how did our best reps handle a similar objection," Qdrant retrieves the most relevant examples in milliseconds, even across thousands of historical conversations.

Omnichannel Ingestion

Modern sales don't happen only on the phone. A typical deal might start with a call, continue in WhatsApp, involve email follow-ups, and be tracked in a CRM. Analyzing calls alone misses critical context.

Syntora ingests from:

Channel Data Captured
Phone calls Full transcription + emotional tone analysis
WhatsApp Message content + response times + media sharing
Telegram Message content + response times
CRM (Bitrix24, amoCRM) Deal stage changes, notes, task completion, pipeline movement

All channels are unified into a single deal timeline, so the AI sees the complete picture — not isolated fragments.

CRM Integration

Supported: Bitrix24, amoCRM (more planned)

What Syntora Reads from CRM

  • Deal pipeline and stage progression
  • Contact and company information
  • Manager assignments and reassignments
  • Task creation and completion
  • Notes and comments
  • Deal amounts and forecasted values

What Syntora Does NOT Write to CRM

Syntora operates in read-only mode on your CRM. We do not:

  • Create or modify deals
  • Change pipeline stages
  • Add tasks or notes
  • Alter any existing data

This is a deliberate choice — we provide recommendations, your team executes them. This keeps your CRM data clean and under your full control.

Sync Approach

Integration works via API with periodic synchronization. New data is pulled regularly to keep the deal timeline current. The connection is encrypted end-to-end (security details).

AI Roleplay Simulator

When the weekly action plan identifies a skill gap, reps can practice in the AI Roleplay Simulator before facing real customers.

Three Training Modes

1. Voice Training (24/7) The rep chooses a scenario — for example, "Cold call: getting past the gatekeeper in a dental supply company." The AI plays the customer role with realistic personality: it can be skeptical, busy, rude, or interested, just like real prospects. Training happens via voice or text, any time of day.

2. Exam Mode The system tests product knowledge and script adherence. The sales manager receives a report: "Ivanov scored 85%, growth area: handling price objections." This turns subjective skill assessment into measurable data.

3. HR Interview Trainer For companies using the Full Management plan, the simulator helps evaluate recruiter skills and train interview techniques — ensuring you hire the right people in the first place.

How Scenarios Are Generated

Scenarios are not generic. They are built from:

  • Real objections found in that company's actual sales calls
  • Industry-specific context (dental supplies, real estate, SaaS — each has different buyer behavior)
  • Individual rep weaknesses identified by the weekly analysis

This means a rep practicing "handling price objections" gets scenarios based on real objections their actual customers raise, not textbook examples.

Infrastructure

  • Hosting: Yandex Cloud (certified data centers in Russia)
  • Data residency: All data processed and stored within Russia
  • Encryption: AES-256 at rest, TLS 1.3 in transit
  • Compliance: 152-FL (Russian personal data law)

For detailed security information, see Security & Compliance.


Back to Documentation | How It Works | Security | syntora.ru