AI / Automation
AI Sales Automation
An automated qualification and outreach engine that scores inbound leads, enriches them and routes each to the right owner within minutes.
- Python
- OpenAI
- n8n
- PostgreSQL
- +1
View Case Study
AI / Voice / Multi-Agent
VeriVoice
AI Multi-Agent Voice Platform
An intelligent multi-agent voice platform designed to automate business conversations and workflows.
- React
- Next.js
- Node.js
- AI
- Voice AI
Overview
Project Overview
VeriVoice is an intelligent voice automation platform that enables businesses to deploy specialised AI agents for customer conversations, qualification and task execution. Rather than one general-purpose bot, each agent owns a narrow remit — reception, qualification, booking, follow-up — and hands off to the next with full context. Every call is transcribed, scored and written back to the customer's CRM, and any caller can reach a human by asking.
Services provided
- AI Agents
- AI Voice Agents
- CRM & API Integrations
- Client
- Confidential — multi-location services group
- Industry
- Professional Services
- Timeline
- 14 weeks
- Year
- 2025
- Platforms
- Web application, Telephony
- Team
- 2 engineers, 1 designer, 1 delivery lead
Challenge
The Challenge
The client was losing qualified enquiries outside office hours and burning senior time on calls that never needed a specialist. Earlier voice bots had been abandoned because callers found them slow and evasive.
Roughly a third of inbound calls arrived outside staffed hours and went to voicemail, where most were never returned.
Previous systems had latency long enough that callers spoke over the agent and the conversation collapsed.
Call outcomes were logged manually and inconsistently, so pipeline data could not be trusted.
Any automated system had to hand over to a person instantly on request, without the caller repeating themselves.
Solution
What We Built
We built a multi-agent architecture where each agent has one job, a small set of tools and an explicit escalation path — then optimised the audio pipeline until conversations felt natural.
Split the conversation into discrete agent roles with typed handoff contracts carrying full context.
Streamed speech recognition and synthesis end to end, with barge-in detection so callers can interrupt naturally.
Gave each agent narrow, permissioned tool access — calendar, CRM, knowledge base — rather than open-ended system control.
Built an evaluation suite of recorded scenarios so prompt and model changes are measured against regressions before release.
Wrote structured outcomes back to the CRM on every call, removing manual logging entirely.
Capabilities
Key Features
What VeriVoice does in day-to-day use.
Specialised agent roster
Reception, qualification, booking and follow-up agents each with their own instructions, tools and success criteria.
Context-preserving handoff
Agents transfer the caller along with the full conversation state, so nobody is asked the same question twice.
Sub-second turn-taking
Streaming speech recognition and synthesis tuned so the agent responds at conversational speed and yields when interrupted.
Live calendar booking
Availability is read and slots are held in real time during the call, with confirmation sent before hanging up.
Human escalation
Warm transfer to an available person on request, on low confidence, or on any topic flagged as out of scope.
Call analytics
Transcripts, outcome tagging, sentiment and per-agent quality scoring in one reviewable dashboard.
Stack
Technology Stack
The tools this system runs on, grouped by the role they play.
Frontend
- Next.js
- React
- TypeScript
- Tailwind CSS
Backend
- Node.js
- Express
- WebSockets
AI & Voice
- OpenAI
- Claude
- ElevenLabs
- Whisper
Telephony
- Twilio Programmable Voice
- SIP Trunking
Data
- PostgreSQL
- Redis
Infrastructure
- Docker
- Vercel
- GitHub Actions
Architecture
How It Works
The path a single request takes through the system, end to end.
INBOUND01
Call received
Twilio routes the call into the platform and opens a bidirectional audio stream.
ROUTE02
Reception agent
Identifies the caller, establishes intent and selects the specialist agent to hand to.
REASON03
Specialist agent
Runs its objective — qualification, booking or support — using its permitted tools.
ACT04
System actions
Calendar slots are held, CRM records created and confirmations dispatched mid-call.
CLOSE05
Outcome & handoff
Transcript, outcome and recording are written back, or the caller is warm-transferred to a person.
Outcome
Results
What changed for the business after launch.
24/7
Call coverage
Every inbound call answered, including nights and weekends.
~0.9s
Median response latency
Measured from end of caller speech to first audio out.
68%
Calls resolved without a human
Across booking, qualification and routine enquiries.
100%
Calls logged automatically
Structured outcomes written to the CRM with no manual entry.
Interface
A Closer Look
Screens from the delivered system.
Agent roster — each role carries its own tools and escalation rules. 
Live call view with streaming transcript and confidence signals. 
Outcome analytics — resolution rate and quality scoring per agent.
More work
Related Projects
Other systems built on similar foundations.
SaaS / AI
Smart CRM
A multi-tenant CRM with AI assistance built into the record, summarising history and surfacing the next action for every account.
- Next.js
- TypeScript
- PostgreSQL
- OpenAI
- +1
View Case Study
Next step
Want a System Like This?
Tell us what your business needs to automate and we'll map out a practical build.
hello@devrox.comWe reply to every enquiry within one business day.