Experience

Applied Data Scientist II, Future AGI (Nov 2024 — Present)

Bengaluru, Karnataka

Led the research and development of the futureagi SDK enabling seamless developer communication with platform features, and contributed to comprehensive SDK documentation.
Developed an MCP server from scratch named futureagi-mcp-server to interact with futureagi functionalities using natural language.
Built Django APIs with designers and frontend teams for seamless integration and meeting product requirements.
Researched on LLM as a Judge and set up a benchmarking pipeline to evaluate the LLM as a Judge based evaluation agents using ground truth data.

Gurgaon, Haryana

Finetuned LayoutLMForTokenClassification language model using transformers library to extract named entities in experience section which uses both text and spatial information from resume obtaining overall f1 score of 0.844.
Researched on efficient single GPU training methodologies using PyTorch and presented it to the entire team of 12 data scientists, useful for faster and cost-effective finetuning of LLM’s.
Extracted bullets with weak action verbs from resume and leveraged semantic search using multi-qa-mpnet-base-dot-v1 contextual embeddings to suggest relevant strong action verbs with the help of OpenSearch (search engine).
Built a scalable event-driven service using AWS SNS and GPT-4o (LLM) to extract and structure information from image PDFs, providing feedback using prompt engineering.

Gurgaon, Haryana

Researched and implemented ASPOSE, a third-party .NET-based tool, to provide support for DOCX Format for parsing resumes uploaded by students across 260+ universities and 10000+ benchmarks.
Retrained Partially Labelled Latent Dirichlet Allocation (PLDA) to predict resume super-sections based on n-grams, improving top-1 accuracy from 81.9% to 89.6% and top-2 accuracy from 92.2% to 96.7% through k-fold cross-validation.
Finetuned bert-base-cased large language model for merging bullets written in multiple lines (NSP), utilizing DVC for efficient data tracking and MlFlow for metrics tracking and obtained an accuracy of around 93%.
Implemented recommendation algorithm that suggests similar jobs and candidates based on users past interactions, enhancing the overall user experience using node2vec in NEO4J and obtained NDCG of 0.76.