The 18-Month Project
I have spent the last year and a half trying to automate myself. Not just my workflows, but my actual face and voice. The goal was simple: create a digital twin that could deliver course modules, explain complex AI concepts, and scale my ability to teach without requiring me to sit in front of a camera for 40 hours a week.
The reality was a lot harder. Getting an AI avatar to not look like a hostage reading a ransom note is a massive challenge. We went through multiple platforms before landing on HeyGen for final production, and even then, the iteration cycle was brutal. This is a work in progress. I am documenting it honestly.
Where the build stands today:
- Voice Cloning: Nailed. The custom voice ID is locked in and consistent.
- Visual Fidelity: Excellent. Three approved avatar IDs that look like me.
- The Glitches: Still fighting them. Language cadence gets robotic on long scripts, and there are occasional scene glitches where micro-expressions fail.
- The Verdict: Not perfect yet, but getting close. Good enough for production with multiple takes.
The Progression - Chronological
These are in order. You can hear and see the evolution. The first video was recorded in January 2024. If you know me, you will immediately hear what is off. That is the point. This is what the process actually looks like.
The Very First One
This is where it started. January 2024. You can hear and see everything that is wrong - the cadence is off, the voice does not sound like me, and the expressions are stiff. This is the baseline. I am sharing it because the gap between this and where we are now is the whole story.
Ryan on the GSX 750
A little later than the first render. I used a topic I know deeply - motorcycles - to test whether the avatar could carry a natural conversation. You can still hear the voice issues clearly. The cadence breaks on longer sentences and the pacing feels mechanical. But the visual match was improving.
Avatar V1 - Platform Testing
Testing the HeyGen baseline after switching platforms. The visual match is significantly stronger here. The voice clone is getting closer. Micro-expressions and blinking patterns still need tuning but the gap from January 2024 is clear.
AI: Experimenting vs Operating
Testing the avatar on real educational content. The script pipeline was not fully built yet at this stage - I was still manually writing scripts and feeding them in. You can hear where the delivery breaks on technical phrasing. This is what drove the decision to build a dedicated script-writing agent.
The Brand Audit Game
First real production test for Ready, Plan, Grow course content. The voice cloning holds up much better on shorter, punchy sentences. Long-form delivery still exposes cadence issues but the overall quality is getting close to usable.
AI Strategy: The Foundation
Testing technical topic delivery. The script agent was generating content in my voice at this point, which made a noticeable difference. When the script sounds like how I actually talk, the avatar delivery improves significantly.
Organizing the AI Workforce
Longer, more complex topic. The goal is to eventually have the orchestrator agent trigger these video generations automatically when new course material is needed. We are not there yet, but this test showed the pipeline can handle complex subject matter.
The Matrix Video
Not everything has to be a production test. This one was just for fun - a Matrix-themed render to see what the avatar could do with a bit of creative direction. I am a Matrix geek, so of course this had to happen at some point. Sometimes you build because you can.
How the Pipeline Actually Works
The avatar is just the final output layer. The real work is the pipeline behind it. When I need a new educational module, the process chains together automatically: a script-writing agent rewrites the curriculum in my exact speaking voice using my brand guidelines and tone, then that script is passed to the HeyGen API using my locked-in voice ID and one of three approved visual avatar IDs.
The system writes the script as me, voices it as me, and renders it as me. The goal is to have the orchestrator trigger the whole chain without me touching it. We are close. Not there yet. But close.
Connect with me
Stay curious, my AI friend. It's the secret sauce - think like you are seven. - Ryan