sumit kumar

Architected a multi-agent LangGraph system that analyzes code quality, security vulnerabilities, dependency risks, and git activity in parallel. Implemented agent orchestration with LangGraph state machines and generated structured health reports with severity-ranked findings and recommendations.

Developed a hybrid deep learning model combining Swin Transformer visual semantics with forensic noise analysis to classify AI-generated vs. real images. Achieved 98% classification accuracy on benchmarks including Midjourney v6 and DALL-E 3, and deployed a live interactive demo on HuggingFace Spaces with a Streamlit UI.

Built a production-style RAG pipeline supporting 5+ document formats (PDF, Word, email, OCR images) from ingestion to semantic indexing with Pinecone. Integrated Llama-3.3-70B via Groq API (~2s latency), exposed FastAPI endpoints for query processing, and generated clause-level JSON evidence with confidence scoring.

Achieved top 800 global rank across 4+ ML competitions using deep learning and Vision Transformer architectures. Built robust training pipelines with 5-fold cross-validation and Optuna hyperparameter tuning, improving accuracy by 12%.

Led development of a multimodal AI pricing system, achieving top 1100 rank (Top 4%) out of 25,000+ global participants. Reduced inference latency by 27% using an ensemble strategy (LightGBM/XGBoost) and decreased RMSE by 14.8%.

Built a real-time pose estimation system using MediaPipe to track 33 keypoints at 30 FPS, providing instant audio corrective feedback for exercise form analysis. Improved movement smoothing and geometric angle calculations across repetitions using Kalman filtering.

Activities and societies: Relevant coursework: Data Structures, Algorithm Analysis, Deep Learning, NLP, Large Language Models, Retrieval Systems, DBMS, Operating Systems, MLOps.

BTech in Computer Science & Engineering (Artificial Intelligence focus) at IIITDM Chennai (Aug 2023–May 2027), covering core CS and advanced AI/LLM systems.