Hire Kunal S. - MLOps Engineer

* AI-Powered Conversational Therapy Tool for Aphasia * Building a novel web application using Streamlit to generate personalized practice scripts for persons with aphasia, leveraging the OpenAI API, GPT-4o-mini * Implemented advanced prompt engineering techniques with DSPy, including fine-tuning, RAG, and bootstrapping * Collaborated closely with 2 Speech-Language Pathologists and conducted user testing with 8 persons with aphasia, iteratively refining the application’s usability and effectiveness * Scalable Markerless Motion Analysis Framework Development * Created multi-camera acquisition system used to collect data from over 700 participants over the past 2 years * Optimized video acquisition software by introducing multithreading, reducing CPU load by 20%, and expanding acquisition capability from 2 to 12 high-bandwidth camera (1Gbps each) * Dockerized video acquisition software, reducing setup time by ~90% and enabling deployment in 4 hospital spaces, 3 external institutions, and as a mobile system * Designed a module using OpenCV to automatically track and annotate research participants in 100+ clinical videos, evaluating 8 different tracking algorithms

* Streamlining data input method by consolidating code and removing manual processes for video processing pipeline using Jupyter, Google Firebase, DataJoint, and other custom Python packages * Improving video collection process from 4 GigE cameras by updating the existing user interface and reducing write-to-disk time by almost 100% using OpenCV and multithreading * Created data processing pipeline to wrangle raw ambulation data (>70 GB collected from 22 mechanical sensor channels, at a sampling rate of 1000 Hz, from 6 able-bodied subjects and 9 transfemoral amputee subjects, across two powered prosthetic legs) * Deployed and maintained intent recognition algorithm and classifier models on a Linux-based embedded controller to make ambulation decisions in real-time for able-bodied and amputee subjects * Developed suite of Python scripts from scratch to pool, sample, analyze, and visualize processed ambulation data to evaluate offline model performance prior to online testing * Performed feature selection analysis in Python using regularization and tree-based methods to determine most relevant features for lower limb pattern recognition classifier * Delivered an offline adaptive algorithm for the lower limb pattern recognition classifier that learns as new ambulation data is introduced to the original classifier model * Designed robust wiring harnesses for new prosthetic leg systems, involving planning the wire routes through existing hardware, implementing the harnesses, and creating documentation for future use * Trained new team members in lab standards for creating and documenting wiring harnesses, 3D printing, software implementation and documentation, and Git

* Validated augmented reality application using Unity3D and C# to aid in pattern recognition training for upper limb amputees and provided hardware and software technical support to a team of clinicians evaluating the application with a series of research subjects * Executed electrical bring-up for 2 powered prosthetic legs, involving design/creation of wiring harnesses and preparation of circuit boards required for the powered leg systems * Built clinician friendly interface and control system for prosthetic ankle using Java (Android) and MicroPython * Developed 3 modular Python packages to perform data processing, training, and real-time classification * Researched the underlying algorithms for a lower limb pattern recognition classifier to aid in the development of a robust and efficient system * Supervised timeline and workload for myself and a second engineer for this project