Building scalable AI systems with real-world impact.

🎓 3rd-Year B.Tech – Computer Science (AI & ML Specialization)
🏅 Herbalife Scholar '25  |  EcoHack Runner-Up '26  |  SheFi Scholar (Season 16)
🔬 Research Intern – NIT Bhopal (2025)

• Building production-grade AI systems across Computer Vision and LLMs
• Developed a Vision Transformer (ViT) foundation model on Indian satellite imagery (Sentinel-2) achieving 93.6% accuracy
• Engineered RAG pipelines with hybrid retrieval, reranking, and hallucination detection for reliable AI systems
• Solved 300+ DSA problems, strengthening core problem-solving and system thinking
• Interested in scalable ML, LLMs, GenAI, Geospatial and real-world AI deployment (MLOps)

📫 kumarisnehlata2005@gmail.com

AI-powered legal document intelligence with ML risk classification, SHAP explainability & RAG Q&A

• Built a full-stack application (FastAPI + React) to analyze legal PDFs — extracting clauses using spaCy NLP + regex and classifying each clause as HIGH / MEDIUM / LOW risk
• Trained a Logistic Regression + Random Forest ensemble classifier achieving 89.6% accuracy and a +30.9% F1 improvement over keyword baseline
• Integrated SHAP word-level feature attribution to explain which terms drive each risk prediction, making legal AI decisions transparent and auditable
• Powered a RAG-based Q&A pipeline (Groq LLaMA3 + FAISS) enabling users to query documents with grounded answers and source citations
• Ships with rate limiting, input validation, PDF report export, and Docker-based deployment

Python FastAPI React spaCy Scikit-Learn SHAP Groq LLaMA3 FAISS Docker

Production-grade RAG system for academic document Q&A with grounded answers, confidence scoring & hallucination detection

• Architected a hybrid retrieval pipeline combining FAISS semantic search (BAAI/bge-small-en-v1.5) and BM25 keyword search, reranked with a cross-encoder (ms-marco-MiniLM-L-6-v2)
• Integrated Mistral-7B-Instruct for answer generation with built-in hallucination risk assessment, confidence labels (HIGH / MEDIUM / LOW), and per-answer source citations
• Supports multi-document comparison queries; exposes evaluation metrics (groundedness, retrieval F1, context utilization) in debug mode
• Deployed on HuggingFace Spaces via Docker; exposes a clean REST API (FastAPI) paired with a Streamlit UI

Python FastAPI Streamlit FAISS BM25 Mistral-7B HuggingFace Cross-Encoder Reranking Docker

End-to-end MLOps pipeline for real-time stock price prediction and trading signal generation

• Built a full production ML pipeline: data ingestion → feature engineering → model training → deployment with experiment tracking via MLflow
• Engineered real-time trading signal generation from live market feeds with automated retraining workflows
• Implemented model versioning, drift detection, and performance monitoring for sustained reliability in production
• Containerized with Docker and structured for CI/CD-ready pipeline deployment

Python MLflow Scikit-Learn Pandas FastAPI Docker

Thank you for visiting my profile.
⭐ Open to AI internships, research collaborations, and innovative ML projects.