Aakash Madabhushi

Built an end-to-end LeRobot v3 data pipeline for SO-101 pick-and-place imitation learning, synchronizing 6-DoF joint state/action trajectories with dual 640×480 RGB streams at 30 FPS and publishing 50 teleoperated episodes (49,633 frames) to Hugging Face. Developed pre-training quality gates over Parquet metadata and H.264 video, validating episode/frame alignment, video decodability and resolution, and NaN/Inf-free action and proprioceptive tensors. Engineered a reproducible PyTorch/CUDA training pipeline for 52M-parameter Action Chunking Transformer policies with configurable augmentation, Weights & Biases tracking, 5K-step checkpointing, and resume support; resolved Windows video-loader and checkpoint failures to complete 30K- and 60K-step runs. Built a resumable physical-robot evaluation pipeline for a 100-trial fixed/randomized cube-placement protocol, adding automatic return-to-start, stale-cache and empty-episode guards, placement-error measurement, and success/grasp/failure logging; completed all 50 fixed-position rollouts. Building physics-based manipulation simulations in Isaac Sim and MuJoCo for SO-ARM101 and Franka Emika Panda arms, validating control algorithms pre-deployment.

Built production Python pipelines on Databricks processing 5M+ rows daily to power analytics and recommendation features across all platform customers. Prevented 14 data incidents before reaching model training by adding automated Delta Lake quality checks to every pipeline. Optimized 75+ customer dashboards by profiling Spark jobs, rewriting slow SQL, and caching reused results, cutting p95 load time from 12.4s to 7.1s (43% improvement).

Built an end-to-end service delivering orthodontists realistic after-treatment smile previews from single phone photos by training a CycleGAN on unmatched before/after datasets. Increased usable patient photos by 20% across 1,500+ real captures by assembling unpaired training sets and adding OpenCV preprocessing to screen out or correct dark, off-angle, and partially blocked shots.

Collaborated with student research collective to collect, analyze, and compile data from Twitter for water-related natural disaster insights. Data Acquisition & Preprocessing: Automated the collection of tweets using the Tweepy API, focusing on keywords related to water-related disasters. Employed text preprocessing techniques including tokenization, stop word removal, and stemming, and leveraged FastText to generate high-quality word embeddings from the preprocessed text data. Model Training & Evaluation: Utilized the Support Vector Machine (SVM) algorithm within the scikit-learn library to discern tweets related to water-related disaster events such as flash floods, tsunamis, droughts, etc. Trained the SVM classifier on a labeled dataset of tweets, applying cross-validation to ensure model generalizability. The model's performance was evaluated using accuracy, precision, recall, and F1-score, achieving an impressive accuracy of 85%. Key improvements were using GridSearchCV for hyperparameter optimization (i.e., learning rate, kernel), which improved the accuracy by 7%. Containerization: Streamlined workflow by leveraging Visual Studio IDE within a Docker Container, fostering a standardized Linux & Miniconda environment.