Implemented contrastive activation steering and activation-space intervention during inference. Accepted for PAKDD Special Track on LLMs for Data Science.

Curated homework problem recommendation system using a Rewriter–Retriever–Reranker NLP pipeline. Ranked relevant practice problems from user queries, using GPT-2, T5 and FLAN-T5 with Hugging Face Transformers in Python.

Developed a CNN model to classify 36 ASL gestures (A–Z, 0–9) with 94% accuracy, improving accessibility for the deaf community. Designed a 3-block Sequential CNN with ReLU activation, dropout regularization and Adam optimizer with categorical cross-entropy loss. Used EarlyStopping and ReduceLROnPlateau to balance class performance; achieved macro-averaged F1-score of 95%.

Built an automated traffic violation detection system for night-time conditions using computer vision and ML. Detected stop-line violations and identified vehicles through license plate recognition using MySQL Connector, PIL, OpenCV and pytesseract.

Achieved ≥95% accuracy on chest X-ray disease dataset for differentiating between corona and pneumonia of 84,818 images within 3 hours and <5ms inference latency using PySpark. Implemented separate validation and test sets; deployed the model via GitHub.

Improved detection of irregular GST transactions using XGBoost on highly imbalanced data. Organized by GST Gov. of India.

AI system to detect and challenge self-deception using linear probes, symbolic sanity checks, and debate agents. Blended AI approach using gpt-oss-120b and gemini-2.5-flash for deception risk scoring and counter-narratives.

Performed EDA and visualization on 1450 news transcripts from YouTube (Euro News) and other sources to study dynamics of repeated news. Used Hugging Face libraries for sentiment analysis, topic modeling and text summarization; visualized results with charts and graphs.