Prakash Bisht

University of Arizona

School of Information, University of Arizona, Tucson, Arizona

Master of Science in Data Science

Sharda University

School of Engineering & Technology, Sharda University, Uttar Pradesh, India

Bachelor of Technology in Mechanical Engineering

Machine Learning Engineer

Indiana University Health, Indianapolis, IN

• Developed a two-part deep learning solution using 2D CNNs to predict Cardiac Ablation Sites from ECG data, achieving 95% accuracy in PVC Identification and 74% accuracy in site prediction. This reduced ventricle mapping time by 1 hour per procedure, increased daily procedure capacity by 50%, enhanced patient outcomes, and generated an additional $28K in revenue per surgery, while saving up to $30,800 in manual labor costs.
• Implemented a distributed DL model predicting hospital length of stay with 87.3% accuracy, saving $55M annually by enhancing resource management. Ingested and processed 940M data points from Google Cloud Storage into BigQuery, using TensorFlow-io on PySpark clusters with GPUs on DataProc for training, and integrated MLflow for model tracking, streamlining patient throughput and care..
• Implemented a Retrieval Augmented Generation (RAG) pipeline using Phi-3-Mini model, converting thousands of pharmacy contracts into a vector store. Automated the extraction of critical business information for decision making and vendor negotiation by fine-tuning a FLAN-T5 model for QnA task, reducing manual review time by hundred of hours, enhancing efficiency, and improving contract management and negotiation processes.

Machine Learning Intern

Lightsense Technology, Inc., Tucson, AZ

• Developed a robust fact-checking automation pipeline using OpenAI’s Large Language Models alongside FAISS (Facebook AI Similarity Search) library to validate LLMs generated outputs, minimizing erroneous outputs by 20%, by devising custom NLP models and techniques to identify and mitigate the hallucinating tendency of LLMs in producing unverified information.
• Fostered the potential of LLMs in scientific research, and anticipated 4 times increase in grant writing procedure efficiency owing to LLMs superior Natural Language Understanding, text summarizing, reason based probing & data generation capabilities, by exploring and implementing their use in spectroscopic technology and research grant development.
• Utilized classic ML techniques to uncover latent patterns, explore anomalies in dataset encompassing findings from various Spectroscopic technologies such as UV-Vis absorbance, emissions & scattering and Fourier Transform Infrared spectroscopy with an objective to advance efforts to address fast diagnosis for variety of medical conditions using ML in spectroscopy..

NLP Summer Research Intern

Deep Target NLP Research Group, Tucson, AZ

• Improved precision by 15% over baseline models by fine-tuning and fusing BioBERT and ClinicalBERT models for Electronic Health Records on ASD. Enhanced this by addressing data imbalance, leading to a boost in precision and recall across diverse label distributions and an increased token representation quality for EHR texts.
• Circumvented GPU memory limitations using gradient accumulation and summarization, combined with a dynamic learning rate scheduler for SGD, resulting in 15% faster convergence. Further streamlined data processing, improving throughput by 25% and reduced inference time by 20% ensuring robust real-time response.
• Delved deep into transformer model behaviors, revealing insights leading to a 10% precision boost on edge cases, while deploying techniques such as knowledge distillation to ensure swift real-time responses in production environments.
• Spearheaded initiatives that reduced clinicians' diagnosis time by 50% by collaborating cross-functionally and integrating insights from data engineering, UI/UX teams, and expert clinicians.

Data Scientist

Embibe, Bangalore, India

• Implemented Deep Item Response Theory (D-IRT) data-driven models via scalable deep learning frameworks that gauged latent traits of 8 lakh students & million questions at once dropping annotation efforts by over 90 percent, as measured by the increased efficiency and accuracy of the data analysis, by utilising Neural Architecture Search (NAS) for model search, custom data generators for handling very large datasets, and exporting models to ONNX for model interoperability.
• Engineered “Model Distillation Techniques” to interpret blackbox DL models and applied them to D-IRT models, resulting in the successful extraction of crucial item characteristics such as 'Difficulty' and ‘Discrimination’, and validated the effectiveness through extensive experimentation.
• Built a unified “Automatic Test Generation Utility” for faculties to generate thousands of tests on-demand & validate before commissioning.
• Collaborated with 6 members agile team to develop features and REST APIs for ATG in a timely manner.

ML Engineer

mlsquare, Bangalore, India

• Developed a python package to facilitate Interoperablity amongst different ML frameworks- (Sklearn-Tensorflow, Surpriselib-Pytorch. etc)
• Implemented an sklearn compatible interface for SVD via an appropriate DL model to support interoperable machine learning.
• Built a NAS interface for AutoML.
• Designed and Implemented a surpriselib compatible interface for collaborative filtering via Deep Factorisation Machines.
• Analysed multiple public datasets in Bayesian workflow using Pyro & contributing chapters to “Probabilistic Programming with Pyro”.
• Investigated meta & curriculum learning strategies to immune training regime to predictor outliers, fostering resource- efficient machine learning.

ML Engineer

IcarusNova Discovery Pvt Ltd, Bangalore, India

Graduate Research Assistant (NLP)

Eller College of Management, Tucson, AZ

Leveraging Natural Language Understanding to develop NLP models for “Autism Spectrum Disorder (ASD) Risk Assessment for Early Diagnosis” by analysing the semantics of Electronic Health Records (NIH-NIMH Grant).

• Implemented a boilerplate wrapper which was capable of loading several individual label Bi-GRU models, test dataset passed in different formats and output predictions from ensemble of models.
• Developed a multilabel model, which outperformed the Ensemble of 7 individual models by reducing the prediction time by factor 130, the outcome was a result of employing more vectorised operations thus optimizing the preprocessing phase significantly.
• Experimented with various adversarial retraining configurations for built multilabel model and benchmarked performance against ensemble of individual models as follows, Few of the following modifications to the model configuration were done to pose adversaries to the model and make it more robust:
• GloVe 42B word vector used in embedding instead of GloVe 6B.
• Synonym Replacement Attack: a method to generate adversarial examples by replacing words in the input text with their synonyms. This approach can potentially confuse the classifier by maintaining the overall meaning of the sentence while changing specific words by utilizing ‘WordNet’ to find synonyms.
• Random Character Swap Attack: It generates adversarial examples by randomly swapping characters in the input text. This method aims to introduce noise into the text, making it more challenging for the classifier to recognize patterns. It swaps a specified number of character pairs to introduce noise into the text.
• Oversampling: It is used to balance the training dataset by oversampling the underrepresented classes.
• The final multilabel model trained using adversarial retraining in combination with oversampling was robust to adversarial samples and gave significant improvement in precision for class labels which had low representation in training dataset therefore poor performance.