How Tellyo answers business calls in Québec French and English with a voice AI agent
Tellyo is an early-stage voice AI company that wants every business call answered, in Québec French or English, without a wait. OneNine designed and built its platform: a real-time phone agent that answers common questions, switches language mid-call and transfers callers to the right person. It went from first commit to a working production setup in about five weeks.
At a glance
~5 weeks
from first commit to production
2
languages, Québec French and English, switched mid-call
1 file
of configuration per client: voice, hours, FAQs, departments
Canada
hosting, with transcripts anonymized before storage
The problem
Many calls to a business don't need a person: opening hours, directions, "who should I talk to?". They still tie up staff, and callers who wait too long hang up and call someone else.
Repetitive calls
Opening hours, service questions and routing calls create wait times and pull staff away from the conversations that bring in revenue.
Québec French is hard
Many voice solutions stumble on local accents and expressions, or respond too slowly to feel like a real conversation.
Privacy and scale
Calls contain personal information, so the platform had to meet Québec privacy expectations, and adding a new business could not mean new development.
How we got there
Tellyo needed a product, not a demo, and one that could serve many businesses from day one.
October 2025
A platform, not a one-off bot
We set up one codebase for the voice agent, the management API and the documentation, designed from the start so that each client is configuration, not code.
November 2025
Real conversations in two languages
A voice pipeline over LiveKit and Twilio, instructions written for Québec French, language detection during the call and transfers to the right department.
November 2025
Reliability first
If a speech, AI or voice service slows down, the agent switches automatically from the real-time model to a backup pipeline, so callers never hit dead air.
November 2025
Privacy built in
Recordings stored per agent on Canadian infrastructure, transcripts anonymized in French and English with Microsoft Presidio, and recording disclosed at the start of the call.
January 2026
Faster and more natural
Sessions moved to a real-time voice model for lower latency, with the backup pipeline kept in place.
The solution
One configurable runtime answers calls for every client. What changes from one business to the next is a configuration file.
Answers in real time
Calls arrive through Twilio and are answered by a natural voice, fast enough to feel like a conversation.
Québec French and English
The agent detects the caller's language and switches during the call, with instructions written for Québec French.
The right department, first time
When a caller needs a person, the agent transfers the call directly to the right department.
A new client is a configuration file
Personality, hours, FAQs, languages and departments live in one file per client, and documented runbooks cover provisioning phone numbers.
Under the hood
A Python voice agent and a TypeScript management API, deployed with Terraform on Azure Canada Central.
How it works, step by step
- 1
Call in
Twilio routes the call to LiveKit; voice activity detection and noise cancellation clean up the audio.
- 2
Understand and reply
A real-time voice model listens and answers, with the caller's language detected during the call.
- 3
Fallback
If the real-time connection check fails, the agent switches to a backup chain of separate speech recognition, language and voice models.
- 4
Act
Tools answer from the client's knowledge, transfer the call to a department or end the call.
- 5
Record
The recording is stored in Azure Blob Storage, organized by agent to keep clients isolated.
- 6
Anonymize
Transcripts are anonymized with Microsoft Presidio, in French and English, before being stored.
Technology
Engineering safeguards
- Automatic fallback from the real-time model to a backup speech pipeline
- Data hosted in Canada (Azure Canada Central), with 30-day database backups
- Transcripts anonymized before storage, recordings isolated per agent
- Call recording disclosed to callers
- Automated tests, including a check that French instructions stay Québécois
- Infrastructure as code and documented runbooks for every operation
Benefits
Time saved
Routine calls are handled without anyone picking up.
- Common questions answered automatically, around the clock
- Callers routed to the right department instead of being passed around
- New businesses onboarded through configuration, not development
Revenue impact
Every answered call is a customer who did not hang up.
- No more missed calls after hours or when the lines are busy
- Staff time freed for the calls that turn into sales
- A multi-client platform Tellyo can sell to many businesses from one codebase
Why OneNine
Why Tellyo built its voice platform with OneNine
Production in weeks, not quarters
About five weeks from first commit to a working production setup.
Reliability by design
We plan for AI services failing, and build the fallback before it is needed.
Québec French and Québec privacy
We build for local accents and Loi 25 expectations from day one, not as an afterthought.
Infrastructure you can trust
Canadian hosting, infrastructure as code and documentation the team can operate.