Multi-Agent AI · Production
01
archon-ai
Production-grade multi-agent research system with a streaming
REST API. Orchestrates Research (ReAct), RAG, Synthesis, and
Critic agents over typed shared state using
LangGraph. Powered by Anthropic Claude
(Sonnet + Opus), ChromaDB for vector retrieval, Tavily for live
web search, all served behind FastAPI.
LangGraph
Anthropic Claude
ChromaDB
Tavily
FastAPI
Python
DL Research · Benchmark
02
ArchPlexBench
From-scratch PyTorch benchmark training five sequence
architectures — decoder-only Transformer, encoder-decoder
Transformer, LSTM, Mamba, and multi-head Mamba — at ~6M params on
WikiText-2. Mamba wins with test perplexity
214.6 (~3× better than the Transformer); LSTM
beats both Transformers in this small-data regime.
PyTorch
Mamba
Transformers
LSTM
Research
FullStack-StockVision
End-to-end stock-forecasting platform fusing Python, C++, and
React. Trains and benchmarks seven architectures
(LSTM, Transformer, CNN-LSTM, GRU, BiLSTM, TCN, linear baseline),
exports to ONNX, then serves real-time inference from a
lightweight C++ HTTP server. Works on any yfinance ticker.
Python
C++
React
ONNX
TensorFlow
Computer Vision · Deep Learning
04
VideoFrameRateEnhancer
Frame-interpolation pipeline that doubles any video's frame rate.
Ships four interchangeable architectures —
U-Net, Diffusion, Transformer, and Mamba — for predicting
intermediate frames. Supports custom training or pre-trained
checkpoints, with optimized GPU memory flow for large-scale video
data.
PyTorch
Diffusion
U-Net
Mamba
GPU Optim.
SimpleRayTracer
Ray-tracing engine built using only the C++ Standard Library.
Implements vector ops, ray-object intersection with spatial
partitioning, and a perspective camera with adjustable FOV,
depth-of-field, and focus — producing realistic global
illumination, shadows, reflections, and transparency.
C++
CMake
Linear Algebra
Graphics
