Ayush Khaneja

▪ Engineer 200k+ unstructured clinical text records from MIMIC-IV data to create artificially labelled datasets, eliminating annotation costs. ▪ Fine-tune and refine smaller LLMs (Llama, Mistral) for private, resource-efficient inferencing with performance like models 8-10x in size. ▪ Implement custom chunking and RAG techniques to further enhance LLM efficiency. ▪ Build and deploy an AWS-based LLM pipeline using Python, Transformers, HuggingFace, and vLLM to structure medical notes, generate precise summaries, and detect mentions of diagnoses, reducing manual effort by over 100 seconds per note

▪ Developed vector-based predictive models using Python, Scpacy, NLTK and TensorFlow to predict diagnosis occurrence with 84-88% F1 score. ▪ Deployed LLMs locally on AWS EC2 instances to extract actionable diagnostic insights from unstructured clinical notes for predictive analysis.

• Delivered financial analysis through SAS programming on client financial data to aid the Audit team in their external audit processes • Developed reusable Tableau visualizations and dashboards from scratch to present relevant findings and insights from audit and other financial data • Managed and maintained a Robotic Process Automation pipeline using UiPath, to automate time-consuming client-related tasks, saving 48-72 hours in manual work per week

• Assisted in the development of an automated pipeline to handle client data, allowing the firm to divert valuable man hours (48-72 hours) to other processes

• Conducted journal entry testing procedures in SAS to prevent management override of control, improving reliability of financial statements • Performed data completeness and data quality checks on large quantities of data

Worked on highly unbalanced data, came up with distance-based approaches to classify data with high bias and implemented a Siamese neural network architecture. Research was presented at ICPRAM 2020.

Flying drones using Raspberry Pi and computer vision.