SYSTEM · v3 · MULTI-REGION

One engine. Two products. Every channel.

Quallix Audit and Quallix Assist are two decisions taken over the same conversation pipeline. Same ingestion, same understanding layer, same tool-calling runtime — different policy and different outputs.

Talk to our solutions team Read the architecture brief

PIPELINE · TURN-LEVEL

Channels in, decisions out.

Quallix Audit reads the log Quallix Assist takes the action

CONFIG RECORD · NOT A FORK

A new client is a config record, not a code change.

Onboarding writes a row. Region, language profile, residency and policy limits are data — never branches in the codebase.

Expanding from India to the UAE changes three fields. The state machine, the tool definitions and the escalation logic are identical.

{
"client_id": "nbfc_delhi_ncr_01",
"region": "IN",
"language_profile": ["hi-IN", "en-IN", "pa-IN"],
"data_residency": "ap-south-1",
"policy_limits": { "call_window": "08:00-19:00", "max_contacts_per_week": 3 },
"integrations": ["ozonetel", "salesforce", "freshdesk"]
}

LATENCY · CONFIDENCE

Performance envelope.

Metric	Target	Detail
Voice turn latency	Under 800ms	Speech-in to speech-out, per conversational turn
WhatsApp response	5–10s	Including retrieval and tool calls
Transcription	Streaming	Partial hypotheses available mid-utterance
Confidence ≥ 0.85	Autonomous	Agent acts without human review
Confidence 0.60–0.85	Clarify	Agent asks a disambiguating question
Confidence < 0.60	Escalate	Handover to a human with full context

INTEGRATIONS

We connect to the stack you already run.

Telephony & CPaaS

Ozonetel, Exotel, Twilio, Knowlarity, Asterisk / SIP trunks

CRM

Salesforce, Zoho, HubSpot, LeadSquared

Helpdesk

Freshdesk, Zendesk, Kapture

Field service

ServiceTitan-style FSM, Zoho FSM, in-house dispatch

Messaging

WhatsApp Business Platform, RCS, SMS gateways

Data & BI

S3 / GCS exports, Snowflake, BigQuery, Metabase, Power BI

Custom

REST APIs, signed webhooks, batch recording drops

MODEL STRATEGY

Bring your own model.

Model choice is a routing decision, not an architectural one. Intent classification and routine extraction run on a cost-optimised model sized for the volume of a full call floor.

Policy-bound decisions — disclosure checks, settlement authority, escalation and anything written back to a system of record — are evaluated on frontier models before deployment, and the evaluation set is kept per client.

Enterprise clients can pin a specific provider, region or self-hosted endpoint without any change to the tool definitions or the state machine.

Take this to your architecture review.

We will walk your team through the pipeline, the residency model and the integration surface against your own stack.

Talk to our solutions team Read the architecture brief