Custom Language Model Demo
A decoder-only transformer with 85M parameters, 12 attention heads, and a 1024-token context window. Trained on 100,000+ synthetic customer service conversations using gradient accumulation and cosine annealing on a single NVIDIA A100.
Adjust the generation parameters below—temperature, top-k, top-p, and repetition penalty—to see how each affects output quality. And if you want to try breaking it, go for it.
Try the Interactive Demo
Chat with the language model and adjust generation parameters in real-time.
Start a conversation
Type a message below to chat with the language model
Press Enter to send, Shift+Enter for new line
Model Configuration
Controls randomness. Recommend adjusting this or Top-p, not both.
Number of top tokens to consider for sampling
Nucleus sampling threshold. Recommend adjusting this or Temperature, not both.
Maximum length of generated response
Penalty for repeating tokens (higher = less repetition)
Show response as it generates
Interested in Custom AI Solutions?
From language models to computer vision, I build end-to-end AI systems tailored to your business needs.
How It Works
Three key components powering conversational AI
1) Synthetic Data Factory
100,000+ customer service conversations generated locally using LangChain agents on my RTX 3080, ensuring diversity in tone and topic.
2) Constrained Training
Trained on a single A100 with gradient accumulation and careful memory management. Every GPU hour counted.
3) Streaming Inference
Real-time token streaming via Server-Sent Events. Adjust temperature and sampling to see how generation changes.
Technical Architecture
🤖 Model & Data
- • Training Data: 100K+ synthetic customer service conversations
- • Parameters: ~85M parameters across 12 transformer layers
- • Context Length: 1024 tokens
- • Framework: PyTorch with custom training pipeline
⚡ Production Stack
- • API Framework: Flask REST API with CORS support
- • Deployment: Docker containerized for CPU and GPU
- • Session Management: In-memory conversation tracking
- • Frontend: React with TypeScript and TailwindCSS