• Developed a RAG system using GPT4 and milqus vector database to generate code for various APIs such as ParcelLab, Stripe and NEWSAPI. Using Similarity search allowed the Documentation to be used as context and thus lead to code generation without fine tuning model• Worked on a LLM system to generate collections of products using GPT4.• Worked on a RAG system to build a recommendation generator using GPT4 and cosine similarity search.• Worked on various LLMs such as LLAMA2, CodeT5, GPT4 and DeepSeek Coder.• Worked on various action detection models for warehouse activity detection. Trained Slow Fast and VideoMAEv2 for activity detection. Slow Fast was then deployed to production• Worked on Temporal Action Localization task using InternVideo. Used VideoMAEv2 for Masked Image Modelling and generating features for our dataset. The whole pipeline was trained end to end for our dataset.• Pioneered the development of a pipeline aimed at synthetically generating novel samples by integrating original activities after transformation. This innovative approach yielded a noteworthy improvement, resulting in a 2% increase in Mean Average Precision.• Set up another pipeline for synthetic data generation. The pipeline used VIBE to generate SMPL human models mimicking the activity. The humans were modified with shape, clothes and viewpoints. The final action video was generated using blender.

• Directed the creation of an internal search engine FAQ generation system, leading a team of PhD-level colleagues and employing technologies like LangChain, Llama2, Hugging Face models, and Vector DB.• Applied and optimized open-source Large Language Models to produce accurate FAQ pairs from health articles, significantly enhancing model performance for healthcare applications.• Achieved a significant percentage reduction in support line call volume, demonstrating the system's effectiveness in user experience improvement and information accuracy.• Trained NLP machine learning classification models (BERT, T5) in TensorFlow to flag 15,000+, high risk aneurysmpatients for checkups based on text data from doctors' notes achieving over 92% accuracy.• Evaluated key search engine metrics to track progress towards business KPIs and communicated insights to top leaderswith weekly data visualizations in Tableau and R.• Created SQL queries to sift through registration click stream data, discovered a drop off of 66% related to feature thatwas changed after insights were communicated and resulted in 1m+ additional user signups.• Processed big data set from Hadoop cluster totaling over 2 terabytes to bring in relevant portions into local Python andR Studio environments, cutting down the entire teams query processing times by 43%

Harnessing the power of cutting edge natural language machine learning models to increase cold email outreach conversions by up to 30%!

- Made enhancements to the Workday calendar product that were eventually deployed to production- Worked in an Agile team and participated in daily Scrum meetings- Deployed and tested code using Amazon AWS EC2 virtual machines.

I developed a tutorial platform for teaching kids programming similar to codecademy.com using frontend JS and taught myself Django to use it to build the backend.