Sonal Sinha

MTech graduate from IIT Madras (CGPA: 9.6) and AI consultant at SoftAge AI, with hands-on experience in machine learning, computer vision, and large language models. Proven track record of developing and deploying data-driven solutions across finance, healthcare, aerospace, and enterprise AI domains. Skilled in building end-to-end ML pipelines using tools like LangChain, FastAPI, and TensorFlow. Winner of the AI RE Hackathon 2025, with a top-performing Stacked Ensemble model. Passionate about solving real-world problems through cutting-edge AI innovation.

• IBM certification on Gen AI and prompt engineering (2025): large language models (LLMs), encoding and decoding, prompt engineering, fine-tuning LLMs, deploying generative AI
• IBM certification on RAG systems (2025): retrieval-augmented generation, vector databases, ML pipelines
• IBM certification on artificial intelligence (2025): neural networks, deep learning frameworks (TensorFlow, PyTorch)
• Data Science and Machine Learning Certification – Finlatics (2025): Statistical modeling, feature engineering, supervised/unsupervised learning
• IBM certification on cloud computing fundamentals (2025): cloud architecture, deployment strategies
• Google Professional Data Analytics Certification (2025): data visualization, statistical analysis, SQL
• SQL certification – HackerRank (2025): SQL queries, database optimization
• Diploma in Banking and Finance – Indian Institute of Banking and Finance (2022): Financial modeling, risk assessment
• McKinsey Forward Learner (2025)

• Winner: AI ReHackathon – CMC Vellore and Kaggle (2025)
• Top 5 finalist: Mando AI Hackathon (2025)

Stacked Ensemble Model for Robot Movement Prediction

Built a stacked ensemble (LightGBM + Random Forest + Meta-Learner), optimized via Group K-Fold and Optuna → 1st place at AI RE Hackathon 2025; Youden’s J score of 0.89.

Multimodal Document Q&A System with RAG for Business Application
Built a multi-format Q&A system using LLaMA, Gemini, LangChain (RAG), FAISS, and Pinecone. Integrated NLP-based semantic search, re-ranking, and deployed via FastAPI → Achieved 85% query resolution and reduced response time by 50%.

CNN Vision Model for E-commerce Image Segmentation & Classification
Developed a CNN with transfer learning (ResNet, EfficientNet) for product image segmentation and real-time classification in a mock e-commerce pipeline → 92% segmentation and 90% classification accuracy; 70% training time reduction.

Banking Campaign Prediction with Random Forest
Performed EDA on 45k+ records and built an SMOTE-balanced Random Forest model to predict term deposit subscript→ Boosted accuracy to 88% and improved conversion by 15%.

VR Rehab Software with MARS Robot & Deep Learning
Designed Unity-based VR rehab software integrated with MARS robot and CNN-LSTM model to tailor stroke recovery exercises → Increased patient engagement by 40% and reduced recovery time by 15%.

Streamlit App for Automated EDA & Feature Engineering
Created a web app for automated data profiling and feature engineering from structured/semi-structured data → Reduced EDA time by 60%, improved model performance by 10%.

Machine Learning Model for Crop Productivity on IBM Cloud
Developed an XGBoost model deployed on IBM Cloud, integrating weather and soil and historical data. Optimized with hyperparameter tuning. → Reduced prediction errors by 20%.

Cryogenic Sensor for Aerospace with ML-Based Signal Optimization
Designed and fabricated a U-bent plastic optical fiber sensor for real-time cryogenic liquid level sensing. Integrated ML models (Random Forest, SVM) for signal optimization and validated performance across multiple cryogens  → Achieved 98% detection accuracy and enabled IoT integration for aerospace systems.

• IBM Virtual Internship on Machine Learning and Cloud (July - Aug'25) – Gaining hands-on experience in ML model deployment on cloud platforms
• Google Gen AI Exchange Program (June - Sep'25) – Exploring generative AI applications and collaboration with industry experts.