Playground

24.5M parameter decoder-only transformer trained from scratch. Llama 2 architecture with RoPE, SwiGLU, RMSNorm, and Flash Attention.

Novel ML classification algorithm that starts from data boundaries and works inward to identify cluster centers. Scikit-learn compatible with superior outlier robustness.

Physics-inspired classification using magnetic force simulation. Applies attractive forces to similar class points and repulsive forces to separate different classes.

Novel attention mechanism that generates text by evaluating against dynamic rubrics first, then reattending to failed criteria. Combines Criteria-Masked Parallel Generation (CMPG) with Failure-Weighted Reattention Learning (FWRL).

Educational demonstration of vanishing gradients in RNNs vs LSTMs with real GPU-computed metrics. Delayed Memory Task design forces gradient flow through entire sequence — RNN gradients decay ~500x across 50 timesteps.

Convert Mermaid diagrams into professional architecture diagrams with 123 icons, Konva.js canvas rendering, draggable nodes, LLM-as-Judge quality scoring, and PNG export.

Interactive Transformer architecture visualizer with 33 auto-rendered canvas visualizations — attention heatmaps, LoRA, MoE, RoPE, Flash Attention, speculative decoding, GQA, SWA, RLHF, quantization, PPL, beam search, and more. Enc-Dec / Dec-Only / Enc-Only modes with full Deep Dive tab.

Comprehensive guide to the theory behind every fine-tuning hyperparameter — LoRA rank, AdamW, gradient checkpointing, quantization, and more. Each concept explained with its ML/DL/math domain, plus an 8-stage pipeline flow diagram with PyTorch as the core engine.

Interactive visualizer for single-qubit, CNOT, entanglement, Bloch sphere, and universal gate sets. Click through 7 tabs to explore quantum computing fundamentals.

Deutsch-Jozsa algorithm implementation and quantum circuit experiments using IBM Qiskit framework.

JSON compression for LLM consumption. Achieves 60-85% token reduction using 3-layer compression: schema extraction, dictionary encoding, and equivalence partitioning.

Extends LangGraph for cross-chain node communication. Enables direct calls between different chains, shared state management, and message broadcasting for multi-agent systems.

Python debugging and tracing utility. Provides detailed execution traces, function call logging, and performance profiling for complex Python applications.

Semantic caching for LLM API calls. Cut LLM costs by 20-40% with one decorator. Caches semantically similar prompts using cosine similarity, supporting multiple backends (memory, SQLite, Redis).

Create and activate Python virtual environments with a simple command. Streamlines Python development workflow in VS Code.

Generate Mermaid flowcharts from your code. Visualize code logic and control flow instantly within VS Code.

Summarize code files in your project with advanced glob pattern support. Generate concise overviews of your codebase.

AI-powered PR description generator for local Git changes. Automatically create comprehensive pull request descriptions.

Refactor, explain, or generate unit tests for selected code using a tiny local LLM via Python. No API keys required.

Create folder structures from ASCII tree representations. Paste a tree diagram and instantly generate the directory structure.

Auto-activate a shared Python virtual environment across multiple projects. Perfect for developers using one venv for multiple related projects.

Visualize Python package dependencies from requirements.txt or pyproject.toml as interactive Mermaid flowcharts.

Multi-provider AI chatbot (AWS Bedrock, Gemini, Groq) with note editing, transform prompts, Mermaid diagrams, and glassmorphism UI.

Export notes to Google Docs with Mermaid diagram conversion, code block preservation, callout support, and multiple export formats.

LLM prompt management with tag fingerprinting (50+ tags), AWS Bedrock testing, similarity detection, and health scoring.