Projects
Curupira
ActiveOpen-source quantitative trading research lab
An autonomous AI agent that runs its own quant research cycle — from hypothesis generation through backtesting to signal deployment. Everything is published openly, including the strategy graveyard of failed approaches.
- 31+ strategies tested, with full failure logs published
- Autonomous research pipeline with LLM-driven analysis
- Hurst exponent regime detection, entropy collapse signals, FVG analysis
- $1.30/day LLM trading agent on Hyperliquid
Real-Time ML Video Production Platform
MonksAutonomous sports video production with multi-model inference
6 configurable ML pipelines processing live multi-camera video streams. Chained inference across object detection, pose estimation, player recognition, and segmentation — orchestrated across 4x NVIDIA L40S GPUs managing ~96 GB GPU memory.
- 7+ models: SAM3, RF-DETR, ViTPose, Cosmos 7B, FastVLM, Mistral Nemo 12B
- LLM-powered autonomous video director for camera switching decisions
- Kubernetes GPU deployment with air-gapped support and NVIDIA NIM
Probabilistic Time-Series Modeling
ResearchHidden Markov Models for algorithmic trading
Financial time-series state transition prediction using HMMs. Applied hmmlearn with Pandas and NumPy for regime detection in currency and crypto markets.