Headline
AI & Machine Learning Engineer | Building Accessible & Production-Ready AI Systems
About
Motivated B.E. student specializing in Artificial Intelligence and Machine Learning at Acharya Institute of Technology (CGPA: 8.5). Passionate about designing and deploying intelligent, accessible, and production-ready AI systems. Experienced in building end-to-end applications spanning full-stack development, LLM integration, hallucination-resistant RAG architectures, and reproducible MLOps pipelines using tools like PyTorch, Hugging Face, DVC, and vector databases. Eager to contribute to innovative engineering teams and solve complex, real-world problems
Education
Projects
End-to-End Reproducible MLOps Pipeline & Deployment
Architected a 5-stage reproducible machine learning pipeline (data ingestion, feature scaling, model training, inference, and visualization) orchestrated with DVC for stage caching and single-command reproducibility. Implemented data and model versioning using Git and DVC linked to a DagsHub remote storage, effectively decoupling large binary artifacts from source control. Trained and evaluated a Random Forest classifier, producing automated confusion matrices and feature-importance evaluations. Built and deployed a live interactive Streamlit web dashboard for real-time inference and model performance monitoring.
View details
VeritasRAG – Hallucination-Resistant Retrieval-Augmented Generation System
Architected a production-grade RAG pipeline featuring hierarchical document resolution chunking and hybrid dense+sparse (BM25) search. Implemented BERT-based semantic boundary detection for document splitting and stored chunk embeddings in vector databases for nearest-neighbor retrieval. Built a custom trust-scoring and bias-estimation module with automated provenance tracking and knowledge lineage graphs. Integrated LLM-as-a-judge evaluation frameworks (RAGAS, FEQA) coupled with NLI-based hallucination detection for factual correctness. Designed a multi-layered defensive gateway to detect and block unsafe or inaccurate outputs prior to delivery.
View details
Project 1: AccessiRead
Built a full-stack assistive web application designed to help dyslexic users read and comprehend complex documents via AI summarization and neural text-to-speech. Integrated Maya-1 Neural Audio Transformer via PyTorch/Hugging Face for low-latency emotional voice synthesis, managing SNAC tokens and Base64 WAV delivery. Engineered prompt pipelines on Groq Cloud to instruct Llama 3.1/3.2 models to generate simplified, cognitive-load-optimized summaries. Built a Flask REST API supporting multi-modal inputs (text, image/PDF via OCR.space), automatic language detection, and neural translation. Implemented an accessibility-first UI featuring OpenDyslexic font, adjustable character/line spacing, and high-contrast themes
View details