Atharva Bhairam

Performed structured annotation and QA review of 500+ AI-generated outputs, identifying quality issues, inconsistencies, and edge cases impacting model performance Categorized outputs and documented error patterns with clear notes to support training data refinement and annotation guideline improvements Reviewed and validated labeled datasets to ensure consistency, accuracy, and adherence to evaluation standards across contributors Provided detailed corrective feedback on model outputs, contributing to a 22% reduction in observed error rates Collaborated with training and evaluation teams to refine review procedures and improve inter-annotator agreement Maintained rigorous quality standards and clear written documentation while working independently in a remote evaluation workflow

Processed and analyzed 500+ weekly call recordings, evaluating audio quality and transcription accuracy across diverse noise conditions Identified audio artifacts such as background noise, clipping, and alignment issues impacting transcription performance Categorized and labeled audio samples to improve model understanding of real-world audio imperfections Provided structured QA feedback on audio-transcription outputs to refine model accuracy

Implemented an end-to-end Retrieval-Augmented Generation (RAG) conversational system using LLaMA 3, FastAPI, and React, enabling accurate, context-aware responses over a knowledge base exceeding 150 documents. Developed a document ingestion and vectorization workflow supporting PDF and DOCX formats, including text preprocessing, chunking strategies, embedding generation, and vector storage in Qdrant, achieving sub-500ms semantic search latency. Designed a robust backend architecture leveraging PostgreSQL, Firebase Authentication, and RESTful APIs to manage user authentication, session handling, chat history persistence, and secure document uploads. Built a modern, responsive frontend interface using React and Tailwind CSS, delivering a real-time chat experience and implementing Context API–based state management for seamless interaction flows. Conducted extensive validation and testing across diverse query scenarios to evaluate retrieval accuracy, response relevance, and hallucination reduction, ensuring reliable and consistent AI-generated outputs.

As an AI/ML Developer at Codelissian, I worked on building and deploying scalable machine learning and NLP solutions end-to-end. I trained and optimized classification models on datasets exceeding 50,000 samples using PyTorch, achieving a 28% performance improvement through effective feature engineering and hyperparameter tuning. I also developed and launched an NLP-based chatbot powered by Hugging Face models, supporting over 100 user intents and serving more than 1,000 users while reducing response time by 40% compared to rule-based systems. Additionally, I implemented a Retrieval-Augmented Generation (RAG) pipeline using Pinecone and LangChain across 3,000+ documents, reducing information retrieval time from minutes to seconds. To support production workloads, I deployed high-availability inference services using FastAPI, handling 200+ requests per minute with 99.5% uptime and serving over 10,000 daily production queries.