Stephen Karasek Email & Phone Number

Quincy, Ma, Us

Hartford, Ct, Us

LLM and NLP Integration:- Developed an in-house LLM with RAG queries for lightning claim prediction.- Trained RoBERTa-based BERT model for lightning keyword recognition, comparing with LLM/RAG and custom NLP models.- Utilized Databricks and Azure Servers to run models, transferring data to and from Azure and local storageModel Development:- Created a new training method for Location-Risk-Score models using Distributional Anchor Regression across multiple datasets.- Distributed training on Sagemaker, using Data Parallelism for large CAE datasets, and Model Parallelism for larger models unable to fit in-memory on single GPU- Built ensemble NLP models in PyTorch combining text, numerical, categorical, and image data to predict lightning-related claims.Data Processing and Optimization:- Pre-processed claims data with normalization, encoding, and oversampling using LLM-generated examples.- Tuned hyperparameters with grid search, Optuna, and Bayesian optimization, applying advanced regularization and K-Fold Cross Validation.Specialized Models:- Designed Bayesian regression to link power quality factors with claims.- Built a transformer model to classify power usage profiles.Data Engineering:- Automated data transfer from AWS S3 to Azure Blob Storage.Evaluation Metrics:- Measured performance using confusion matrices, ROC curves, and learning rate scheduling.

Quincy, Massachusetts, Us

LLMs (text, image & multimodal):- Built a microservice system for automated documentation of CAE simulation results with LLMs using a RAG-based approach.- Fine-tuned LLMs (LLaMA, Mistral, Hermes) for subject matter expertise.- Handled 3D object detection with Mask R-CNNs, SIFT, and vision-based LLMs like LLaVA.Cloud Infrastructure & Automation:- Managed keyword extraction, data visualization, and formatting with AWS services (Lambda, S3, ECS, EKS, EC2, SageMaker).- Developed cloud infrastructure with Python, Go, and Docker, achieving 60% cost savings over prior systems.Simulation & SaaS Development:- Led AI-guided simulation environment development using tools like OpenFoam, SU2, and PyTorch, reducing workflows by 90%.- Built SaaS for CAD-based virtual testing, cutting product development time and cost by 75%.Data Engineering & Model Training:- Created a pipeline to train ML/NN models on 200,000+ n(3+1)D data blocks, linking design features to physics simulations.- Devised self-correcting neural network algorithms to prevent error propagation.Rapid Prototyping & ML Applications:- Designed and evaluated ML prototypes for unsupervised analysis of CAD models and simulation results.- Parallelized neural networks and physics models for exponential cloud-based speedup, and implemented with PySparkBusiness Impact:- Secured $310,000 in grant funding, and over $100,000 in other assistance, diversifying offerings and expanding the business.- Developed business relationships with dozens of customers, closing over $100k of revenue from software sales and contracts

Cambridge, Ma, Us

• Architected and developed supervised machine learning models to analyze acoustic and optical time series signals using. Methods used included VARMAX, ARIMA and LSTMs.• Built & tested CNN model for speed and accuracy on satellite images• Simulation and model based design for system execution, prediction of behavior and viability- Algorithms developed for imaging & sensing, optical systems, guidance, and navigation- Parallelization of algorithms, processes and code for drastically reduced run-time; Parallel algorithms written in MATLAB, C++, C, CUDA and Simulink.- Helped design compute clusters running parallel code, including NVidia GPGPUs (both Kepler and Maxwell-based architecture), and Intel coprocessors (Intel Xeon Phi family)• Optical Semiconductors & Sensing Systems- Modified and implemented hybrid finite-difference & beam propagation method, to determine input signal changes from linear, linear dispersive and nonlinear effects

Boston, Massachusetts, Us

• Designed FDTD (finite difference time domain) simulation in MATLAB, MEX and C to predict optical response of target, based on perturbative, nonperturbative or strong field laser source. Determined which imaging strategy was best for a target, and reducing cost of techniques• Developed linear & polynomial regression models for analyzing simulation datasets• Numerical calculation of interactions between laser source and electrons in atomic and molecular structures, ranging from simple square wells to Hydrogen ions, yielded system resonance conditions and response• Linear and nonlinear effects arose from time-dependent external sources (continuous and pulsed), whose conditions were varied over multiple simulations. With respect to the system resonance conditions, the source conditions ranged from matched, to mismatched and severely mismatched• Adapted and devised gauge symmetry conditions for laser-matter interactions, in order to simplify calculations for laser sources, boundary conditions, finite difference methods, and increased accuracy of operator splitting• Finite Difference schemes used were the explicit-staggered 2nd FDTD-Q and 4th order GFDTD-Q. Boundary conditions analyzed included convolution (exact absorbing), Dirichlet and Neumann.• Imaginary Time Propagation (ITP) and Fourier Analysis methods revealed Energy States, indicated momentum & intensity; further filtering showed approximate population of energy states at a given time• Configured MATLAB to use GPGPU cores, using NVidia's CUDA, for parallel processing. Developed algorithms to run in parallel, to exponentially reduce run time. Redesigned simulations to make use of new resources• Repaired & realigned KECK 3D fusion microscope, fine-tuning optics and reconfiguring electronics. Wrote LabVIEW VIs based on instruments’ pre-existing C# code to facilitate most of the hardware communication

Lead engineer for project to develop NIR laser sensor system from off-the-shelf components- Lasers & sensors continuously scan tubes of varying material, filled with different liquids, for air bubbles- System is modular in design, in order to be easily integrated with any of the different machines, materials and liquids that customers might possibly wish to use- Performed cost benefit analysis of several off-the-shelf solutions, vs building a system in-house- Designed supporting electronics, including laser driver- Programmed microcontroller to regulate laser diodes and sensors, reporting pertinent data back to larger machine the sensor system is attached to- Presented project updates regularly to company president and top engineering management

Natick, Ma, Us

• Programmatic GUI manipulation in MATLAB, Java, and Java emulation within MATLAB• Lead architect for tool consisting of several GUIs and APIs to automate image collection for documentation. Tool allowed writers to run desired automation scripts, compare new images with document controlled copies, and interact with command line and file system to check out, replace, then submit them• Designed automation scripts and generated structure of scripts based on requests in the form of XML documents. Upon execution, at specified intervals, a frame grabber function would capture the results.• Collaborated with Developers, Quality Engineers and Technical Writers to improve processes and user examples

Pleasanton, Ca, Us

• Member of a team that developed a LVAD (Left Ventricular Assist Device) system• Assisted with design of watchdog circuit for the device• Generated requirements and test protocols for product reliability• Developed electronic test fixtures to ascertain feasibility of design• Performed electrical functional testing to determine design reliability and durability• Designed modular LabVIEW programs to emulate and interact with hardware, simulating communication and field scenarios. Programs were distributed for development cycle within company, subsidiaries, and to partner companies• Analyzed intensity of Optical Curing System, determined where losses and temperature deviations occurred along path length and their likely causes (such as UV divergence when using appropriate filter)

Boston, Ma, Us

• Designed and modified MATLAB scripts and algorithms for optical instruments• Modified polarization system in microscope to reduce losses• Modeled system in Zemax-EE• Researching methods for improving existing OQM system and analysis techniques, using different kinds of polarized light and different techniques of polarization

Woburn, Ma, Us

SHORT TERM CONTRACT• Designed multi-threaded program in Parallax’s SPIN/Assembly RTOS for use by Propeller microcontroller, using SPIN, Assembly and C• Program allocated resources and cores to running functions of flying car, such as navigation and GPS, fuel and flight indicators, and ran constant debug mode to ensure proper functionality• Performed analysis, debugging and assistance with redesign of custom motherboard; board used SPI interface to communicate with different subsystems

Lexington, Ma, Us

• Designed object oriented MATLAB GUI and script to automate data collection for infrared detectors; script controlled a black body source and oscilloscope• Modified MATLAB device drivers for the black body source and oscilloscope• Developed radiometry equations and MATLAB scripts for processing data into readable graphs• Programmed C++ code to build libraries for use with MATLAB• Designed cold-stop for detector aperture in SolidWorks

• Tested and validated the circuitry and mechanics of Pressurized Liquid Extraction systems• Installed, connected and maintained office network with Windows Server 2003 and Windows XP• Wired, installed and networked the security cameras

Lawrence, Massachusetts, Us

• Projects included validation of a custom developed software tool, a learning management software suite, a water purification system, an electronic prescriptions submittal system, and a dialysis therapy machine• Designed and executed test scripts and performed engineering studies for design, development, and manufacturing of a medical device system used for dialysis therapy• Performed R&D studies, including data collection and analysis• Developed proficiencies in test authoring, test execution and requirement development, as well as good documentation practices and working in a regulated industry (FDA)• Received bonus and certificate of achievement from Vice President after validation of electronic prescriptions submittal system; worked two 60 hour weeks to finish on schedule

Master’S Degree, Electromagnetics & Optical Engineering

Bs, Electrical Engineering

Education record