Turning raw CRM leads into qualified, sales-ready conversations — automatically.
A proof-of-concept outbound voice agent that dials new CRM leads, holds natural Hindi/Hinglish qualification calls, and writes budget, location, BHK, and timeline back into existing CRM workflows — built on LiveKit, Sarvam AI, and Google Gemini with TRAI/DLT and DPDP-aligned compliance.
Lead generation was fine — qualification could not keep up.
A real estate developer’s bottleneck was the first conversation. Manual telecalling queues let leads go cold, notes stayed inconsistent, and outbound calling carried real TRAI/DLT and DPDP exposure.
Leads go cold in hours
Queue-bound telecalling delayed first contact and burned conversion potential.
Language and tone expectations
Leads expected natural Hindi/Hinglish — not rigid English scripts or IVR menus.
Inconsistent qualification notes
Sales needed clean budget, location, BHK, and timeline fields — not messy call notes.
Compliance and call reliability
Outbound dialing required consent windows, suppression checks, and precise outcome classification for rings, voicemail, and failures.
Client
Real estate developer (confidential)
Industry
Real estate lead qualification
Integrations
LiveKit · Sarvam · Gemini · n8n
Engagement
Compliance-first voice AI POC
What we learned before we designed anything
Discovery focused on speed-to-lead, Hindi/Hinglish conversation quality, regulatory constraints, and call-edge-case reliability — not just LLM dialogue.
Natural Hindi/Hinglish wins
Leads respond best to conversational speech, not scripted or robotic IVR tone.
Sales needs structured fields
Budget, location, BHK, and timeline must be filterable CRM data, not freeform notes.
Compliance is non-optional
TRAI/DLT and DPDP rules had to gate every dial — consent, suppression, and calling windows.
Call classification matters
Rejected rings, voicemail, no-answer, and failures each need distinct CRM outcomes.
Fit the existing stack
The AI layer had to extend Zoho/HubSpot via n8n — not replace the client’s automation.
Monolithic dial + talk services
Combining compliance decisions and conversation in one service made audits and scale harder.
Workflow-only compliance
Relying solely on n8n checks risked accidental bypass if workflows changed later.
Fully scripted IVR flows
Hardcoded scripts captured data but sounded robotic; open chat alone risked missing fields.
Operations Lead
Telecalling / sales operations · Real estate developer
Goals
- • Qualify every new lead within minutes
- • Free telecallers for sales-ready conversations
- • Stay compliant by default on every dial
Pain Points
- • Expensive call-volume spikes
- • Slipping response times
- • Inconsistent notes and compliance risk
Dispatch decides; the voice worker converses.
A compliance-gated dispatch service triggers LiveKit dials; a dedicated worker runs speech, Gemini reasoning, and structured capture, then webhooks results back into CRM via n8n.
Compliance-first outbound qualification, end to end.
Hybrid scripting, live tool-call capture, and provider adapters turn messy telephony into clean CRM outcomes.
Manual queues decided who got called, when, and whether consent rules were followed.
Gate every dial before it leaves
- ✓Consent, suppression, and calling-window checks live in the FastAPI dispatch service
- ✓New CRM leads trigger automatic dial attempts within minutes — not hours
IVR scripts or inconsistent human notes failed language and data-quality expectations.
Natural Hindi/Hinglish qualification
- ✓Sarvam speech + Gemini reasoning with a fixed compliant greeting and free-flow dialogue after
- ✓Agent stays silent until the call is confirmed answered — never greets ringtone or voicemail
Post-call transcript parsing missed fields and delayed CRM updates.
Confirm details during the call
- ✓Tool calls record budget, location, BHK, and timeline mid-conversation
- ✓Agent verbally confirms details with the lead before closing the session
Failed calls were lumped together and CRM/telephony lock-in limited options.
Reliable classification, swappable vendors
- ✓Distinct outcomes for connect, no-answer, voicemail, busy, and technical failure
- ✓Telephony and CRM providers stay configuration — Exotel/Plivo, Zoho/HubSpot via n8n
Built for Indian outbound voice AI reality.
LiveKit for real-time voice, Sarvam for Hindi/Hinglish speech, Gemini for dialogue, Silero VAD for turns, and FastAPI + n8n for compliant dispatch into existing CRMs.
Separate dial decisions from conversation ownership.
We centralised compliance in dispatch, kept the worker focused on talking, and designed every call to produce a structured CRM result with no manual step.
Constraint mapping
Documented speed-to-lead, language, TRAI/DLT and DPDP exposure, SIP unreliability, and CRM/n8n fit.
Compliance-first dispatch
Placed consent and calling-window gates in the dispatch service so workflows could not bypass them.
Hybrid conversation model
Fixed opening line for brand/compliance, then LLM tool-calling dialogue for natural qualification.
Answer discipline & outcomes
Held speech until confirmed answer and classified transient rejection signals without tearing sessions down early.
Provider abstraction
Wrapped telephony and CRM behind adapters so vendor swaps stayed configuration changes.
What happens on every outbound lead
What the POC was designed to deliver
Minutes to first contact (projected) - Design target: replace hours of queue-dependent dialing with automatic dispatch after CRM entry.
Consistent qualification fields - Every completed call yields the same structured budget, location, BHK, and timeline data.
Telecallers freed for sales-ready work - First-touch screening shifts to the agent so humans focus on qualified conversations.
Compliance coverage by default - Consent and suppression checks run on every dispatch attempt before any dial.
Architecture proven end to end - POC showed an LLM voice agent can hold coherent Hindi/Hinglish calls and update CRM workflows cleanly.
Provider-agnostic foundation - Telephony and CRM adapters keep the client free to change vendors without rewriting agent logic.
Challenges & Learnings
Rejection vs hang-up signals
Transient telephony rejection events looked like disconnects — keeping sessions alive until confirmed state was critical.
One worker per call
Isolating conversations prevented overlapping voices and blocked slow calls from stalling others.
Async webhook resilience
Retries ensured brief downstream outages never silently dropped completed results.
Compliance in dispatch
Putting consent and windows in the dial layer made the system easier to audit than conversation-only checks.
Edge cases define voice AI
Reliability broke on rings, voicemails, and disconnects — not on the happy-path dialogue.
Where the platform goes from here
The architecture proves an LLM-driven voice agent can hold a coherent, compliant, Hindi/Hinglish qualification conversation and deliver clean data straight into existing CRM workflows.
“This engagement was delivered as a proof of concept for a real estate prospect. Client identity is protected, and Results figures are projected design targets — not measured production metrics. The recommended next step is a scoped, metered pilot.”