FinSights: Production-Grade Financial Intelligence System
Hybrid RAG over SEC 10-K filings: structured KPI lookup plus semantic search across 614K sentence embeddings, streamed from AWS Fargate at $0.014 per query with answers cited to exact filing sentences.
Text-to-Pose: CLIP-Conditioned Diffusion for 3D Human Pose
CLIP-conditioned diffusion model that generates 3D human skeletons from captions. A retrieval evaluator reproduces the published R-Precision of 0.797, and a conditioning analysis looks at how CLIP handles left and right.
SocrAItic Circle: Multi-Agent LLM Debate & Evaluation
Multi-agent LLM debate framework with judges built on argumentation theory, rebuilt as a local LangGraph and MCP system. Hybrid retrieval grounds the evidence, and failure-mode instrumentation caught nine silent defects.
Multi-View 3D Scene Analysis
10k+ line pipeline for pose-guided multi-view feature correspondence on ETH3D, with RANSAC validation, occlusion-aware multi-view tracks and a COLMAP mapper integration.
Protein Structure Prediction
HMM, CRF and BiLSTM models for protein secondary structure prediction on the CB513 benchmark. The CRF reached 67% accuracy using evolutionary features and specialized β-sheet features.
Vision-Language Late-Interaction Retrieval
ColBERT-style retrieval that pairs ViT patch tokens with CLIP text tokens through a learned projection, fine-tuned with LoRA on 1,500 Flickr8k pairs. MaxSim scoring with symmetric InfoNCE cut the training loss by 38%.
Artist Classification
Comparison of SVM-SIFT-BoVW, autoencoders and CNNs for artistic style classification across 50 artists. The classical SVM reached 89.2% accuracy, ahead of ResNet-50 and VGG-16.