Michael Trehan

You want a character that answers whatever the player says. Build the loop in order: text, game state, voice, then a face.

You are choosing or budgeting for an AI video tool. Compare how each one charges, then work out what a finished minute costs your team.

You have a clip to make and a blank prompt box. Copy a prompt that fits the job, then learn the order of parts that lets you write the next one yourself.

Your callers mix two languages in one sentence and the agent trips. Fix it stage by stage, then prove it with a test set you can rerun.

You have a Flutter app and a voice agent. Set up WebRTC on both platforms, then put a live avatar on screen without writing a signaling server.

You have a voice agent and a front end. Give the agent a live face with one token endpoint, one hook or service, and one video element.

You have one shoot day and a script that needs three locations. Split the shot list first, then film so the generated shots cut in cleanly.

The media is encrypted before you write a line of code. Your work is deciding who gets a token, what the browser can see and where an embed may run.

You need live audio between a user and an agent. Decide whether a managed room or your own socket carries it, and what each one leaves you to build.

One media server is full and you need a second. Split signaling from media, pin each session to a server, and scale without cutting live calls.

You need a voice agent that answers fast and sounds right. Pick the architecture by what you must control, then put a face on it.

Your agent feels slow and you cannot say which stage is to blame. Trace each turn with standard GenAI spans and add the avatar spans the conventions leave out.

Your Pipecat agent answers Spanish and French callers in an English voice. Detect the language per turn and switch voice, prompt and avatar audio with it.

Your Rust service has to put live audio and video in a browser. Pick the crate, write the signaling, and send real media.

Your server renders frames and a browser has to play them live. Put them on a video track, and WebRTC handles encoding, pacing and loss.

You need the right voice on a Vapi assistant, maybe more than one, and a face that keeps up. Pick the voice, set it, then hand off between voices.

Your avatar's mouth has to match the words. Learn the mouth shapes, where to get them, and how to keep them on the audio clock.

Your agent cuts people off, or keeps talking when they interrupt. Both come down to how it detects speech and what it cancels when it does.

Your users hear themselves, or your voice agent keeps interrupting itself. Find out what the browser cancels, what it cannot, and how to fix the route.

Your call or avatar feels slow and you do not know which stage to blame. Measure each one in the browser, then fix the largest first.

Some of your viewers are on fiber and some are on a train. Send three layers, let the SFU pick one per viewer, and check that each layer is really on the wire.

Your call will not connect, or it connects and stays silent. Find the layer that failed in a few minutes, then apply the matching fix.

You are picking a transport for a voice agent or avatar. Decide by what travels on each hop, then wire the two together.

You have an idea nobody can picture yet. A concept trailer shows it in a minute or two, if you make it for the right audience at the right time.

You are deciding how media should flow in a call, maybe one with an AI agent or avatar in it. Count the endpoints, then pick the topology.