I build and study machine learning systems: deep learning models, LLM and retrieval pipelines, and data-driven analysis. I care about how a system behaves, not just whether it runs: where it fails and how to measure it.
Interests: deep learning · NLP and LLMs · retrieval and knowledge graphs · speech and audio · data science
| Project | What it is | Stack |
|---|---|---|
| Global-context narrative reasoning | Checks whether a character backstory is consistent with a full-length novel, using evidence, thematic and temporal memory layers with iterative retrieval. Built on the open-source ComoRAG framework. First Runner-Up, Kharagpur Data Science Hackathon 6.0, IIT Kharagpur. | Python, BGE embeddings, knowledge graphs, GPT-4o-mini |
| Hybrid quantum federated learning | Studies how skewed (non-IID) client data degrades federated training when each client runs a hybrid quantum-classical model, and tests proximal regularisation and entropy-weighted aggregation as fixes. | PyTorch, PennyLane |
| Audio deepfake detection | LSTM + Transformer-encoder classifier for real vs synthetic speech, trained on 68,889 samples with MFCC, chroma and spectral features. Part of a three-model comparison. | TensorFlow/Keras, librosa, scikit-learn |
| Retail pharmacy analytics | IIT Madras capstone on first-hand data from a retail pharmacy: working-capital analysis, cash-flow forecasting, and safety-stock and reorder-point models for 20 medicines. | Statistics, forecasting, Excel |
| Learning collapse detection | Streaming pipeline that flags declining student engagement from clickstream data. Big Data Analytics course project. | Spark, Python, Next.js |
Applied AI Engineer Intern, Jurident AI. Worked on a legal research platform: multilingual support across 8 Indian languages, multi-document retrieval, and testing for hallucinated precedents.
