Workflow Automation Objective: Automate after-call workflow graph creation using a chatbot interface and guided-path graph summarization using LLMs Ensured structured output generation from LLMs using Pydantic class definitions, allowing data validation & serialization Leveraged prompt engineering, function calling & query re-asking through Instructor package to generate workflows from queries Modified open-source packages for compatibility, dockerizing and deploying the tools to reduce workflow creation time by over 60%

Reimagining Tooling as Coding (RTaC) Framework Objective: Create a low-latency tool-use LLM which matches closed-source LLMs in performance while being cost-effective Created RTaC framework to convert tools to python functions, promoting docstring-reading capabilities in coding-base LLMs Utilized PEFT to finetune Code Llama 7B & DeepSeek 1.3B on 2500 examples of manually cleaned GPT generated synthetic data Achieved a competitive performance to GPT-4 at 20% of the cost, supporting dynamic tool addition & mathematical/iterative logic

Maternal and Child Health Monitoring in LMICs Objective: Estimate 6 key indicators of health status in low income countries using numeric features derived from satellite imagery Trained a Boosting on Error ensemble to predict 4 level residuals using Random Forest & XGBoost, with MCRMSE of 10.759 Utilized Dask package on an 8GB dataset for EDA, finally using RF.feature_importances_ parameter to perform feature selection Secured 1st position out of 30 participating teams in a Kaggle competition conducted by the lab as part of the internship program

Lunar Turing-GAN Super-Resolution Model Objective: Create a high-resolution map of the Moon using a pipeline of Image Super-Resolution models Proposed a novel GAN-based architecture with separate adversaries for ensuring accurate reconstruction of craters and hills Developed a Lunar Atlas by correcting coordinates & stitching together individual image patches from Chadrayaan-2 TMC payload Created a pipeline capable of tiling and super-resolving an image using Lunar T-GAN, HAT, RealESRGAN and sharpening algorithms Achieved a competitive SSIM of 0.794 while increasing spatial resolution from 5m/pixel to 30 cm/pixel, a 16x magnification Report link: report ...

Semantic Segmentation of Remote Sensing Images Objective: Develop an e!cient CNN architecture using Tensorflow for semantic segmentation of geospatial images into 4 classes Improved standard U-Net architecture by adding dilated & depthwise separable convolutions achieving a mean IoU of 0.73 Employed unified focal loss in order to work with class imbalanced Landcover.AI dataset, along with custom dataloader pipeline