Data Scientist
CurrentBackend optimization:•Optimized the backend code repository by refactoring circular imports, removing redundant package calling, and abstracting multiple function calls and class structures.•Time profiling the implemented optimizations resulted in an ~8% decrease in the software cold start time.Bug monitoring:•Monitored and fixed bugs from the log notebooks in an initiative to help the Operations and Customer Success (CS) teams.•Developed a complete strategy for automating customer issue ticketing and reporting using Zendesk, Hubspot, and Slack.Large Language Model (LLM) deployment:•Deployed open-source LLMs over a Llama CPP endpoint. Methods of Quantization and LoRA were applied to increase model performance with lower VRAM usage and finetuning the said models. The finetuning pipeline is currently deployed on Vertex AI. Models deployed – CodeQwen-7B, BGE-base.LLM implementation:•Implemented Google's Vertex AI and Amazon’s Bedrock, specifically the model endpoints for their text and code generation.•Reduced the data payload being sent to the LLM by implementing in-memory context caching to maintain a serverless architecture.Document summarization:•Developed a complete document summarization architecture based on the Retrieval Augmented Generation (RAG) strategy to generate document-based question-answering using LLMs. The system utilized Langchain in tandem with Chroma DB.•Set up the foundational RAG architecture for document summarization using graph databases, Neo4j, and DirectX.•Created subsequent data models for the document summarization process. Big document summarization:•Created a special use case for context-adaptive summarization of very large documents using clustering and LLMs.RAG Optimization:•Improved the document chunk retrieval in RAG architecture using FlashRank, introducing an average improvement in retrieved chunk relevancy by 25%.Client engagement:•Developed a finetuned Mistral model for metadata classification.
