I build machine-learning systems end to end: I put together the data, train and evaluate models, and ship them as working services. My work spans applied ML on real-world data, computer vision, NLP for Indian languages and agentic AI.
M.Tech Data Science, Amrita Vishwa Vidyapeetham, Bengaluru · AWS Certified Cloud Practitioner
| Project | What it is | Result |
|---|---|---|
| AppealIQ | Predicts which health-plan denials get overturned, using 42,710 real California medical-review decisions. Power BI, ML, NLP and anomaly detection. | ROC-AUC 0.736 on unseen years; top-risk decile 90.5% overturned |
| Return-Risk Scorer | Flags cash-on-delivery orders likely to be returned before they ship. CatBoost + SHAP behind FastAPI, with a live demo. | Deployed and tested in CI; explains every score |
| KIC-COCO | First MS COCO-scale Kannada image-caption benchmark (250K captions, 50K images), with six captioning architectures and two zero-shot VLMs evaluated on it. | SacreBLEU-4 68.65, BERTScore 84.02 · paper under review |
| AgroVision-7 | Fine-grained rice-leaf disease classification across 7 classes, using an Optuna-tuned ViT + Swin + ConvNeXt ensemble. | 97.48% accuracy, 0.975 macro-F1 |
| SustainSched | AI agent that schedules Kubernetes workloads by grid carbon intensity. LangGraph + local Llama 3.2 + XGBoost forecasting. | Forecaster R² 0.917 across 14 regions |
| Sentient | Agentic RAG terminal that tracks 8 AI news sources and checks every claim against its sources before publishing. | Draft → verify → revise loop |
More: real-time geospatial monitoring (Kafka, PostGIS, PySpark, PyTorch)
| ML & data science | scikit-learn, XGBoost, CatBoost, SHAP, Optuna, pandas, SQL, Power BI |
| Deep learning & CV | PyTorch, timm, ViT / Swin / ConvNeXt, Grad-CAM |
| NLP & generative AI | Hugging Face Transformers, IndicTrans2, LangGraph, RAG, Ollama |
| Engineering | FastAPI, Docker, Kubernetes, Kafka, PySpark, PostgreSQL / PostGIS, GitHub Actions |