mlx-debate-lab: Multi-Agent LLM Debate on Apple Silicon
LangGraph + MCP debate system running entirely locally on 24GB unified memory against ~34GB of model weights. Nine documented failure modes found building a working reference implementation, no API calls.
OSCD Sentinel-2 Satellite Change Detection
Classical baselines vs. a 269k-parameter Siamese U-Net on Sentinel-2 bitemporal imagery, Apple Silicon/MPS, no CUDA. 1.6x F1 over the classical floor (0.378 vs 0.237), with leakage-free thresholding and four overturned claims documented in the open.
FinSights: Production-Grade Financial Intelligence System
Hybrid RAG architecture with DuckDB stratified sampling and semantic retrieval. Processes 72M→1M sentences achieving $0.017/query cost with exact citation provenance.
Text-to-Pose Diffusion
CLIP-conditioned diffusion model with cross-attention for 3D pose generation. Hybrid CNN-Transformer architecture with anatomical constraint enforcement and kinematic chain validation.
Multi-View 3D Scene Analysis
10k+ LOC pipeline for multi-view scene reconstruction with pose-guided filtering, occlusion handling, and RANSAC validation on ETH3D dataset.
Vision-Language Late-Interaction Retrieval
ColBERT-style ViT+CLIP system with LoRA parameter-efficient fine-tuning. MaxSim token→patch alignment achieving 38% InfoNCE loss reduction and 100% retrieval accuracy on Flickr8k.
Protein Structure Prediction
Implemented HMM, CRF, and BiLSTM architectures for secondary structure prediction. CRF model achieved 67% accuracy on CB513 benchmark using evolutionary features.
SocrAItic Circle
Multi-agent LLM debate framework with multi-phase cycles, iterative refinement, and YAML-driven orchestration for enhanced reasoning and logical consistency evaluation.
Artist Classification
Comparative study of SVM-SIFT-BoVW, CAEs, VAEs, and CNNs for artistic style classification. Classical SVM approach achieved 89% accuracy on 50-class dataset.