02 — Projects
Selected work.
001
In Progress
Allay-sama
Imagine a YouTube livestream with an AI as the host. AI has massively taken over the autonomous side of computers, from simple data algorithms to generative text — the goal is to push this further by creating an AI that gets as close to AGI as possible using today's tech, and proving it can be just as entertaining and charismatic as the average popular streamer.
Powered by a locally-hosted LLaMA 3.1-8B with a custom LoRA adapter, running on GCP with an NVIDIA L4 GPU. Continuous testing and debugging have shaped it into what it is now, and the work isn't stopping until it runs flawlessly.
System Architecture
LLaMA 3.1LoRAGCPNVIDIA L4Discord.pyMongoDBPythonTTS
002
Tooling
Dataset Pipeline
Automated data cleaning and profiling toolkit for preparing training datasets. Built with pandas, AutoClean, and ydata-profiling to handle messy real-world data at scale — deduplication, type inference, outlier removal, and quality reporting.
PythonpandasAutoCleanydata-profiling
003
In Progress
RL Game Agent
Training an AI with reinforcement learning to play games — starting with turn-based and strategic games like Inscryption, Liar's Bar, Buckshot Roulette, Keep Talking and Nobody Explodes, Uno, Monopoly, visual novels, and card games, then scaling up to larger open-world environments like Minecraft. The aim is building agents that can reason about hidden information, bluff, and adapt to unpredictable human opponents.
Training environment
PythonPyTorchReinforcement LearningOpenAI GymStable-Baselines3Computer Vision
shiyo@gpu ~ nvidia-smi --query-gpu=name --format=csv,noheader
NVIDIA L4
shiyo@gpu ~ python -c "from transformers import AutoModelForCausalLM; print('ready')"
ready
shiyo@gpu ~ python serve.py --model allay-sama --port 8080
✓ Model loaded. Inference server running on :8080
# another day, another model served.