Projects
Papyrus
Built a full-stack Retrieval-Augmented Generation (RAG) system for scientific literature Q&A, enabling users to ask natural-language questions and receive cited, grounded answers backed by source documents.Combines dense vector search and keyword retrieval through reciprocal rank fusion, reranks candidates with an LLM relevance scorer, and synthesizes grounded answers with inline citations.
Sentinel
Real-time anomaly detection across vibration, audio, and log streams from industrial equipment. Fuses a VAE on time-series sensors, a CNN on mel-spectrograms, and a text classifier on machine logs into a late-fusion ensemble with SHAP explanations. Kafka ingestion, Triton serving, TimescaleDB storage, and Evidently drift monitoring.
Distill
Fine-tuned a small language model with SFT and DPO to extract structured JSON from scientific papers - authors, methodology, datasets, findings, limitations. Quantized and served with vLLM behind a FastAPI gateway with schema-constrained decoding. Outperforms frontier APIs on accuracy while cutting cost-per-call by an order of magnitude.
Synthesis
A multi-agent system that answers complex research questions across scientific literature. Five specialized agents - planner, retriever, reader, critic, synthesizer - collaborate through structured memory and tool use, with every decision logged to a custom trajectory viewer. Includes a 50-question eval harness with LLM-as-judge scoring and a hallucination-guard pipeline.
Work Experience
- The docstring of _get_train_sampler in GRPOTrainer comment to mention num_iterations > 1 fixed the existing issue.
- Engineered and optimized machine learning architectures by applying foundational mathematical principles (linear algebra, multivariate calculus, and probability), enabling the custom tuning of loss functions and optimizers that improved model accuracy by 15% over standard baselines.
- Architected, trained, and evaluated complex deep learning models using PyTorch and TensorFlow, specifically optimizing forward and backward passes to reduce training times by 30% and maximize computational efficiency on datasets exceeding 500GB.
- Spearheaded the development of an end-to-end AI research pipeline designed to process and synthesize massive, unstructured scientific data, automating the extraction of latent patterns and reducing manual analysis time by 40%.
- Implemented advanced neural network architectures, including graph-based models, directly from research papers, mathematically verifying adjacency matrices to accurately capture non-linear relationships and boost predictive performance by 22% on highly relational data.
- Constructed robust data ingestion and feature engineering workflows, resolving bottlenecks in data preprocessing to decrease data pipeline latency by 45% and accelerate model convergence during training.
- Conducted rigorous model evaluation, hyperparameter optimization, and ablation studies, establishing highly reliable baselines and improving overall inference accuracy by 18% for downstream analytical tasks.
Skills
AI / ML / DL
Frontend
Backend
Database & Cloud
MLOps & Deployment
Languages
Highlights
Achievements
GATE : All India Rank 5472 out of 69242 in Data Science and Artificial Intelligence (DA) branch.
Agents Course : Successfully Completed Agents Course by Hugging Face
Open Source
GitHub Activity
