Solfreaks VoiceAgent builds phone agents that answer incoming calls or place outgoing ones: a clinic reception line, an appointment reminder, a first-pass lead qualification call. The caller speaks, the agent turns the speech into text, a language model decides what to say or do, and a synthetic voice replies.
We'll be direct about where this works. Voice agents do well on narrow jobs with a clear goal and a short list of outcomes. They do poorly at open-ended conversation, and anyone promising a receptionist that handles everything is overselling.
What the line comes with
- A written call flow. The script, plus the few ways each call can end: booked, rescheduled, transferred, callback requested.
- Telephony, connected. A number and a SIP trunk or Twilio-style provider, tied into your existing phone setup where that's possible.
- Live booking and lookups. The agent checks your calendar or booking system and confirms a real slot instead of taking a message.
- Transfer to a person. A warm transfer to a staff number, or a callback ticket if nobody picks up.
- Call summaries in your CRM. Each call ends with a short summary, the transcript and the outcome, sent to your CRM or sheet.
- A test report. Results from test calls using the accents and phrasing your real callers use.
What makes or breaks a call
People forgive a slow web page but not a phone line that goes quiet. We keep delay down by streaming speech recognition, starting the spoken reply before the whole answer is written and using a faster model for simple turns. Callers also interrupt, talk over the agent and say 'hello?' into silence, so the agent has to stop speaking when cut in on and recover gracefully. We tune this on recorded test calls, not on paper.
Language needs case-by-case testing. Urdu, English and the mix of both that many callers speak don't perform equally across speech-recognition and voice providers, and regional accents shift the results. We try several providers on recordings of your callers before choosing one, and we tell you what we find even when it's unflattering.
Recording and consent rules differ by country and sometimes by region. We can make the agent announce that the call is automated and recorded and follow the rules you give us, but we build software and don't give legal advice, so check the requirements with a lawyer. Phone regulations in Pakistan also affect how numbers and call routing can be set up, so we confirm the design with your carrier before promising it.
From brief to live line
- Choose the call. We listen to real recordings, with permission, and pick the one call type that is frequent and predictable.
- Write the flow. Every branch, every handover and every sentence the agent must never say, agreed with you on paper.
- Build a prototype line. Typically two to four weeks gets you a number you can ring and talk to.
- Test with real voices. First our staff, then yours, then a few friendly customers. We fix the misheard words, long pauses and awkward handovers.
- Go live on a slice. Perhaps after-hours calls only, then widen as the transfer and failure rates settle.
Who a phone agent suits
Reception and after-hours cover, appointment confirmations and reminders, lead qualification and simple status checks are the sensible starting points. A follow-up message on WhatsApp after the call often works better than a second call.
A voice agent is the wrong choice if your callers are often upset or the calls need judgement each time. It's also a poor choice at a few calls a day, where a person or a good voicemail costs less. If the calls involve health details, say so early, because it changes how we handle data; our healthcare software page explains that side. For agents that work across email, chat and databases and not just phones, see AI agent development.
What the bill is made of
The quote moves with the number of call types, the integrations, languages, telephony setup and expected volume. Running costs (telephony minutes, speech recognition, the language model and the voice) are billed per minute or per use by the providers, and they add up at high volume, so we work out the cost per call from your numbers before launch. We offer a fixed-scope build for the first agent and a monthly plan for monitoring, call review and tuning.
A narrow agent is usually live in four to eight weeks, and each extra call type adds time. The quote follows a short discovery call.
Why us for phone agents
We're based in Rawalpindi, so we can test with Pakistani accents and mixed Urdu and English from the first week. We're a small team and you talk to the people who build and tune the agent. Publishing apps on Google Play taught us to test on real devices with real users, and we bring that habit to this work. The telephony and provider accounts are in your name. When you're ready, talk to us about your calls and send a few sample recordings if you have them.
Open-source tools and references
A few public projects that come up a lot in this kind of work. They belong to other people and teams. We link to them because they are worth a look, not because anyone pays us to.
- openai/whisperOpen speech recognition model.
- livekit/agentsFramework for real-time voice and video agents.
- pipecat-ai/pipecatOpen-source framework for voice and multimodal assistants.
- rhasspy/piperFast text-to-speech that runs locally.
Frequently asked questions
Will the agent understand Urdu, English and a mix of both?
Sometimes well, sometimes not, which is why we test before promising anything. Speech recognition quality varies by provider, accent, line quality and background noise, and mixed-language sentences are harder than either language alone. We run recordings of your own callers through several providers and show you how often each one gets it wrong. If none is good enough, we'll say so.
What happens when the caller wants a person?
The agent offers a transfer as soon as it's asked, and also when it has failed to understand twice in a row or the caller sounds upset. During staffed hours it transfers to a nominated number with a short summary; out of hours it books a callback and raises a ticket. We test handovers harder than anything else, because a bad one is what callers remember.
Can we record the calls and keep transcripts?
Technically yes, and transcripts are useful for tuning. The rules on recording, and on what you must tell the caller, depend on the countries involved, so the agent can announce that the call is automated and recorded. We build software and don't give legal advice, so please check the rules for your callers' locations. We can also keep summaries only and delete audio on a schedule.
Can the agent make outbound calls for reminders or sales?
Reminders and confirmations to people who already deal with you work well. Cold sales calling is another matter: many countries restrict automated calls and require prior consent, and telephony providers can block numbers that behave like spam. We'll build outbound flows for contacts who agreed to be called and decline projects aimed at lists of people who never did.
How much will each call cost to run?
Every minute involves telephony, speech recognition, a language model and a synthetic voice, each billed by its provider, so cost per call moves with call length and with the providers we pick. Before launch we calculate an expected cost per call from your volume and typical call length, then compare it with the real figures after the first few weeks of live traffic.
Talk it through with us
Tell us what you are building and where it hurts. You get a straight answer, a plan and a price, and if we are not the right people we will say so.
