Cally is earlier work and no longer in active development. This page describes what we built and what I learned from it.
Context and problem
A large part of the Turkish economy still runs on the phone. Clinics, workshops, agencies and local businesses win or lose customers on whether someone picks up. Calls arrive during lunch, after hours, or while every person is busy with another customer. Each missed call is a missed appointment, an unanswered price question, or a support problem left to rot.
The existing tools don't fit. Call centers are priced for enterprises. Chatbots live on websites, but the customer is on the phone. So the phone line, the channel where the money actually moves, stays the least covered part of the business.
What the product did
Cally put an AI voice agent on the business's phone line. The agent answered, listened, and worked out what the caller needed. A booking request became an appointment. A price question got an answer or a qualified handoff to sales. A support issue was captured with the details that matter and routed to the right person.
The business got its calls handled around the clock, plus summaries of what happened on the line. Cally was not a chatbot with a phone number bolted on. It was built around real call traffic: interruptions, background noise, people who change their mind mid-sentence.
My role
Co-founder and product engineer. I worked across product strategy, voice flows and call handling, engineering, onboarding, pilots and sales. In practice that meant designing how a call should go, building the flows, putting the product in front of businesses, and sitting in the onboarding calls where the real feedback lived.
System architecture
- Business phone line
- Any hour, any accent
- Real call conditions
- Listens in real time
- Understands the request
- Identifies intent
- Appointment booking
- Sales questions
- Support issues
- Call summaries
- Handled workflows
- Integrations
A call came into the business line and the agent picked up. While the caller spoke, the agent identified what the call was actually about. That intent drove everything downstream: the call was either handled directly, like booking an appointment, or routed with context to the right person. The business saw a summary instead of a missed-call notification.
Engineering decisions
Intent first, not menu trees
Phone menus ask callers to fit themselves into options. Cally did the opposite: the agent listened first and derived the intent from what the caller says. The trade-off was real: free speech is messier to handle than key presses, and evaluating intent quality took constant work with real call recordings. But it was the difference between a phone robot and something a caller will actually talk to.
Built for the phone line, not the browser
Voice on a phone call is unforgiving. There are no typing indicators and no retry buttons; silence longer than a beat feels broken. Every part of the flow was designed around that constraint, which ruled out approaches that would be acceptable in a chat interface.
Turkish-first from day one
The product served Turkish businesses, so handling spoken Turkish well was not a localization task, it was the core of the product. Flows, intents and test calls were built in Turkish first, not translated after.
Product decisions
Pilots before scale
Cally grew through pilots with real businesses instead of a broad launch. A pilot forced the product to survive one business's real call traffic end to end. That was slower than launching wide, and it was worth it: every pilot rewrote some assumption we had on paper.
Onboarding is part of the product
I ran onboarding personally. Watching a business set up its flows showed exactly where the product confused people, and those calls shaped the setup experience as much as any roadmap discussion.
Sell answered calls, not AI
Businesses don't buy technology descriptions. The pitch was concrete: calls answered, appointments booked, questions handled after hours. The AI was the how, never the what.
What I took from it
Cally taught me how unforgiving real-time AI is once it meets real users, and how much of a product is decided in onboarding calls rather than in code. Both lessons go straight into how I build Derbent. Usage and outcome numbers from the pilots are not published here; this site does not publish estimates.