VoicePeri
From an answered call to a record someone can act on.
From the idea.
Into the real world.
I co-founded VoicePeri in 2026 and built its website, configured and tested the ElevenLabs voice agents, and connected lead generation, CRM and scheduling. The outbound qualification workflow I built in Make.com is still running.
Built the website, configured and tested the ElevenLabs voice agents, and connected the lead generation, CRM and scheduling workflows for launch.
A conversation needs a destination.
Small businesses need calls handled and a structured summary delivered to the places they already work: text, email, a dashboard and their CRM.
Connect every handoff.
I built the marketing site, configured the agents and connected calls and their summaries to CRM and scheduling. I also built the outbound path, carrying a lead from outreach through an AI conversation, qualification, a structured summary, and CRM follow-up.
Keep the platform. Fix the drift.
Longer scenario tests exposed inconsistent tone, pacing and expressiveness. I kept ElevenLabs, moved to a newer conversational model, rewrote the persona and tone rules, tuned voice settings, and reran the same scenarios after meaningful changes.
Checked across the whole conversation.
The voice stayed natural and recognizable across the full scenario set, including longer and interrupted calls. Passing the full scenario set is required before launch; this case does not claim a live client deployment.
Every handoff.
One connected flow.
Explore the moving parts and the work behind each connection.
Conversation
An ElevenLabs voice agent handles the conversation. I configured its persona, tone, and behavior, then tested it across realistic call scenarios.
An overview of the project. No customer or production data.
Change the voice provider, or keep the platform and fix the configuration?
The agent had to sound like one consistent person for the length of a real call, including interruptions and off-script questions, not only in a short demo.
Tone changed during longer conversations.
Expressiveness and pacing went uneven.
Some replies sounded flat and mechanical.
Stretches of a single call felt like slightly different personalities.
Keep the platform, fix the configuration
Move to a newer, more responsive conversational model, rewrite the persona and tone rules, tune the voice configuration, and repeat the scenario tests after each meaningful change.
- A newer, more responsive conversational model.
- A rewritten system prompt with explicit persona and tone rules.
- Tuned voice configuration: stability, similarity, pacing and expressiveness where supported.
- A regression run of the same scenario set after each meaningful change.
I reran the same scenario set after every meaningful change, including long and interrupted calls, until the voice stayed natural and recognizable across an entire conversation. Passing the full set is required before launch.
Connected operations
- 01Understand the operation
- 02Connect the systems
- 03Train the people
- 04Keep things moving