Kiri is my personal AI project. It is a research-driven voice system focused on one thing that is usually ignored: high-quality Khmer speech.
I started Kiri because Khmer doesn't have enough modern speech models. Not enough datasets. Not enough open and serious research. So, I decided to build one myself.
Kiri is built by me with my own time, my own mistakes, and my own money. And to build Kiri, I needed one very special power.
My Superpower
My superpower was not intelligence. It was not talent. It was not even motivation. My real superpower was very simple: Compute.
As long as I had GPUs, I could build Kiri. I could train. I could fail. I could try again. Compute was the engine that turned my ideas into real models.
And today… I am losing that superpower.
I only have one to three days of budget left. After that, I will no longer have enough compute to train new models for Kiri. That is what this story is really about. Not a model, not a release, but the moment when my main power disappears.
The Final Run
Right now, one last model is still running. In about 7 hours, it will finish. This may become the last model I can train for a while. Not because I ran out of ideas, but because I ran out of money for compute.
Even harder to say: After this model is ready, I may not have enough compute to serve inference for everyone. For public testing, I can only afford around 4 hours of total generated audio.
Four hours. For everyone. After that, the server must stop.
The Vision Beyond the Model
This last model is not only my final training run; it is a very small window for people to hear what I have been building. But this model is not my real dream.
My real goal for Kiri has always been speech-to-speech. * Not just text-to-speech.
Not just reading sentences.
Voice → Model → Voice.
I want someone to speak in Khmer, and Kiri listens, understands, and replies—in Khmer—with a real voice. Low latency. Natural rhythm. Real interaction. This last model is only one piece of that future. Only a voice. Only a foundation.
The Cost of Research
When people see Kiri, they say:
“Nice project.” “Nice demo.” “Nice downloads.”
I see something else: Failed runs. Out-of-memory errors. Broken checkpoints. Logs frozen at 2:47 AM. Me, refreshing the terminal, hoping the job didn’t crash.
I didn’t burn money on GPUs because I was confident. I burned money because I didn’t want to stop asking questions. That is why losing compute feels like losing a superpower. Compute gave me speed. It gave me feedback. It gave me the permission to fail fast.
When you lose compute, you don’t just stop training models. You slow down a dream.
It’s Not the End
But I keep reminding myself: The GPU was never the real power. The real power was choosing to build Kiri when nobody asked me to. Choosing to work on Khmer speech when it was easier to ignore it.
So when this run finishes, when the last logs stop scrolling, and when the last free audio sample is generated… I won’t see it as the end.
I’ll see it as a checkpoint.
I may have lost my superpower today. But I did not lose my goal.