AI enthusiast & Data Engineer | M.S. in Artificial Intelligence @ Northeastern University

I like AI/ML research, enterprise-scale data engineering, building systems in deep learning, NLP, and computer vision. Please visit: https://mjsushanth.github.io/

• FinRAG/FinSights: Production-Grade Financial Intelligence System — Hybrid dual-path architecture combining structured queries (DuckDB/SQL dimension tables) with semantic retrieval. Processes 72M→1M sentences via stratified sampling with temporal weighting across regulatory eras. FinSights
• Local LLM debate pipelines using MLX workbench, Local Models, Langgraph, MCP servers. Debate lab
• Text-to-Pose Diffusion: Built a CLIP-conditioned diffusion model with cross-attention + anatomical loss for 3D pose generation.
  • researched concepts on Motion/3D data: (pose representation, N-joint hierarchical mapping, kinematic chains, pelvis-spine-extremity validation) and the architecture of Hybrid CNN-Transformer Diffusion, CLIP, Dual-Pass CFG and Anatomical Constraint Enforcement. See Report here., See Design here.
• Multi-View 3D Scene Analysis: Created a 10k+ LOC pipeline with MV scene analysis, pose-guided filtering, occlusion handling, and RANSAC validation on ETH3D. See Design Flow.
• T2P-Reboot: Rebuilt the text-to-motion project as a measurement-first diagnostics repo: silent decode bugs, probed studies on encoder's directional-word blindness, clip's spatial blindness, showed FID swings 14x on sample size alone, and independently reproduced a published R-Precision benchmark to within 0.87σ on a released checkpoint, with no training of its own. See repo here.
• Satellite Change Detection. Classical methods and Siamese U-Net on OSCD Sentinel-2 bitemporal pairs. Change detection Research
• Protein Structure Prediction: Implemented HMM, CRF, BiLSTM; CRF reached 67% accuracy on CB513 using evolutionary + context features. See Report here.
• SocrAItic Circle: Multi-Agent Debate LLMs workflow, designed with multi-phase debate cycles, iterative refinement, YAML-driven orchestration, and judge modules.
• Artist Classification: Compared SVM-SIFT-BoVW, CAEs, VAEs, and CNNs; SVM achieved 89% accuracy on 50-class dataset.

• Re-implementing transformer enc-dec problem sets, ViT, etc. on Apple MLX, comparing MLX - MPS. Translating Transformer theory into correct code, tensor shapes, masks, residual paths, data structures, and training behavior from a blank page. MLX Practice - Transformers
• Multi-vector ViT+CLIP with LoRA and ColBERT-style MaxSim retrieval Demo Notebook.
• An example workflow of ML-Serving using Gitub CI/CD and AWS Lambda, SAM Infrastructure. Src Code. , Notes here. Notes.
• Usage of Optuna and MLFlow using a synthetic time-series generator Code.

• Do visit my portfolio Portfolio
• Personal: Personal
• Slide decks:
  • FinSights — production-grade financial RAG system, architecture through evaluation. Deck
  • T2P-Reboot — taking a text-to-motion pipeline apart: silent decode bugs, encoder blindness, a reproduced benchmark. Deck
  • MLX Debate Lab — multi-agent debate with retrieval-grounded evidence, judged on a laptop. Deck