Ahmed Ibrahim · Voice AI engineer & consultant

Build a voice agent people can talk to.

I help teams build voice agents and get them working in production. Bring me in to write the code or work through a call flow that keeps going wrong.

Hands-on individual contributor. Scope and availability by email.

A conversation, under the hood
Caller audio
  1. Speech to textHear the words
  2. LLM + conversation stateDecide what happens nextYour data & business tools ↔
  3. Text to speechSay something useful
One common architecture. Turn detection, interruptions, and recovery run across the whole conversation.

Where does your agent need work?

You might have a workflow on paper, a prototype that needs to handle real callers, or a live agent with one stubborn problem. Each is a reasonable place to start.

Build the first version
Define the call flow, connect speech and language models, and wire the agent to the systems it needs to use. Start with a specific task the agent can complete and a clear way to hand off when it cannot.
Fix the awkward conversations
Investigate slow replies, interruptions that get missed, and agents that keep talking after the caller changes direction. Trace the turn through the pipeline before changing the model or prompt.
Make tools and context reliable
Connect the LLM to business APIs and the right information. Handle timeouts, validate actions, and make sure the agent only confirms an outcome after the system has confirmed it.
Get ready for production
Work through deployment, call traces, failure recovery, and evaluation. Include the difficult calls: silence, background noise, unavailable services, and callers who go off script.

A good answer can still be a bad call.

If the agent takes too long, talks over someone, or loses track of a correction, the conversation breaks. These are engineering problems I work on alongside the language model.

What an interruption changes

Illustrative sequence · not measured timing
Stopping the audio is one part of the job. The next response also needs to account for what the caller heard and what they just changed.

The model sits inside a larger system.

A common voice agent uses speech recognition, an LLM, and speech synthesis. Speech-to-speech models combine some of those steps, but you still need to manage tools, interruptions, and failures.

I work with LiveKit and WebRTC for real-time audio, Python and TypeScript for agent behavior, and the services around them. The architecture should fit the call you need to handle.

Real-time audio, with room for interruptions. Speech is a conversation, not a recording.

WebRTC · LiveKit · Speech recognition
Three parts. One conversation. Select a step.
Ahmed Ibrahim smiling outdoors
Ahmed Ibrahim · Senior software engineer

Voice AI has been my day job since 2021.

I work on voice agents at Goodcall. My experience spans three generations of the agent, from Dialogflow to the LLM systems we build today.

My current work includes real-time speech pipelines, model routing and fallback, and the infrastructure behind each conversation. I work across the stack because a failure on a call rarely stays neatly inside one component.

I have been building software since 2016. Before Goodcall, I worked on chatbots, full-stack products, and a software company I co-founded in Cairo.

Read my background
  1. Dialogflow → hybrid agents

    Full Stack Engineer, Goodcall. Worked across Gen 2 and Gen 3 as the agent evolved from NLU toward LLMs.

  2. The fourth generation

    Senior Software Engineer, Goodcall. LiveKit, WebRTC, speech pipelines, and LLM orchestration.

  3. Now

    Working across the call

    Conversation behavior, model routing, fallback, and production systems.

My voice AI work at Goodcall. Dates and roles are also on my About page.

Work with the person writing the code.

You can bring me in for a focused consultation or discuss hands-on implementation with your team. We agree on a useful piece of work and how to judge whether it is done.

  1. Start with the call

    Describe who calls, what they need, and what a successful outcome looks like. If you already have an agent, show me where it gets stuck.

  2. Agree on the work

    We identify the constraints and choose a scope: a technical review, a working call flow, an integration, or a fix to an existing system.

  3. Build, listen, adjust

    Check the behavior against representative calls. Review failures and tradeoffs with your team as the implementation takes shape.

  4. Leave a clear handoff

    The code and decisions should make sense to the team maintaining them. Agree on the documentation, deployment, and follow-up the project needs.

Want to hear a voice agent before we talk? Try the assistant on my homepage. It responds to questions about my work. Your project may need a different architecture and a different voice.

Before you get in touch

Can I hire you to build a voice agent?

Yes. Get in touch about implementation, an architecture review, or a specific problem in an existing agent. I work directly with your team as an individual contributor. Tell me what you need and your timeline so we can discuss scope and availability.

What does a voice AI engineer do?

I connect the parts of a spoken conversation: audio transport, speech recognition, the language model, business tools, and speech generation. I also work on the behavior around them, such as when to respond, how to handle an interruption, and what happens when a service fails.

Can you help with LiveKit and LLM voice agents?

Yes. My work includes LiveKit and WebRTC, Python and TypeScript, speech-to-text and text-to-speech pipelines, and LLM orchestration. That includes context, tool calling, model routing, and fallback behavior. The starting point is your use case and the stack you already have.

Can you improve an agent we already built?

Yes. Send a description of the failure and, if you can share one safely, a redacted call trace. Long pauses, repeated answers, missed interruptions, and unreliable tool calls each need a different investigation. I would start by reproducing the problem and identifying where it occurs.

Should we use a voice platform or build a custom agent?

That depends on the workflow. A managed platform may cover a straightforward phone flow. Custom code becomes useful when you need more control over conversation state, integrations, or how failures are handled. An architecture review can help make that decision before you commit to a build.

How much does a project cost, and when can you start?

We need to agree on the scope first. A review of one call flow is a different job from building and deploying an agent. Email the use case, what already exists, your target date, and any budget constraints. I will confirm availability and discuss an appropriate scope with you.

Tell me what you want the agent to do.

A short email is enough to start. Include the use case, what you have built so far, and when you need help. Share redacted examples if they make the problem easier to explain.

Choose your email app or Gmail. Both open a draft with a few project questions for you to fill in before sending.

me@ahmed-ibrahim.com