SHIVAM KUMAR

Iteratively explore, share, and prep data for the machine learning lifecycle by creating reproducible, editable, and shareable data sets, tables, and visualizations. Iteratively transform, aggregate, and de-duplicate data, and make the data visible and shareable across data teams. Iteratively develop prompts for structured, reliable queries to LLMs. Used popular open source libraries such as Hugging Face Transformers and PyTorch to fine-tune and improve model performance. Model review and governance: Track model and pipeline lineage and versions, and manage those artifacts and transitions through their lifecycle. Model inference and serving: Managed the frequency of model refresh, inference request times and similar production specifics in testing and QA. Used CI/CD tools such as repos and orchestrators to automate the preproduction pipeline. Enable REST API model endpoints, with GPU acceleration. Model monitoring with human feedback: Create model and data monitoring pipelines with alerts both for model drift and for malicious user behavior.

Forecasted the demand for car rentals on an hourly basis having Lightgbm as the best performing model with RMSE of 33.5. Proficiency in crafting robust algorithms by spearheading projects such as the development of center line features and road classification systems. Performed time series forecasting of traffic congestion at different junction based on historical data. Performed analysis using autocorrelation, stationarity, auto regression, ARIMA, SARIMAX and compared with the LSTM and GRU. Worked on Road Classification and Calamity classification that meant to increase the safety of the person by 68% because of intelligent handling of the surrounding. Utilize SOTA model like Vision Transformer in place of the pre-existing models like CNN to increase the performance of existing model by more than 83%. Built Owner-Manual RAG based LLM that simplifies the process of obtaining information about the vehicle, serving as a modern, interactive alternative to the traditional owner’s manual. Migrated the code base to PyTorch to make it run on GPU which make it 24x faster. Utilised CUDA resource, model compression and transfer the file to ONNX and run it on Tensorrt for higher inference performance.