Raj Muchhala
ML EngineerAI is analyzing your overall score…
Identifying your key strengths…
Evaluating your skill match against the job requirements…
Assessing your cultural and operational fit
AI is analyzing your overall score…
Identifying your key strengths…
Evaluating your skill match against the job requirements…
Assessing your cultural and operational fit
Senior ML Engineer @ Workday | Building Enterprise AI Agents That Transform Work
With 10 years of experience building and scaling AI systems, I specialize in architecting and deploying end-to-end machine learning solutions in production. Currently, as a Senior Machine Learning Engineer on Workday’s Agent Factory team, I focus on building and operationalizing agentic AI systems, including LLM-based workflows, tool orchestration, and scalable inference. Previously at Lumen5, I developed LLM-driven recommendation systems and agentic experiences, with a strong emphasis on reliability, performance, and deployment at scale. Earlier in my career at Arya AI, I led teams delivering AI solutions for the banking sector, spanning facial recognition, document intelligence, and signature detection, and co-developed a global API marketplace that significantly contributed to revenue growth. I’m passionate about turning cutting-edge research into robust, real-world AI systems that create measurable business impact
Rochester Institute of Technology
Master's degree, Computer Engineering
January 1, 2014 – January 1, 2016
University of Mumbai
Bachelor's degree, Electronics Engineering
January 1, 2010 – January 1, 2014
University of Mumbai
Bachelor of Engineering, Electronics
N/A – Present
Rochester Institute of Technology
Master of Science, Computer Engineering
N/A – Present
Workday
Senior Machine Learning Engineer
February 1, 2026 – Present
Toronto, Ontario, Canada · Hybrid
Lumen5
Senior Machine Learning Engineer
August 1, 2022 – February 1, 2026
Mississauga, Ontario, Canada
Arya.ai
Senior Research Scientist
August 1, 2021 – August 1, 2022
Mumbai, Maharashtra, India
Arya.ai
Research Scientist
August 1, 2020 – August 1, 2021
Mumbai, Maharashtra, India
Arya AI
Senior Research Scientist
August 1, 2020 – August 1, 2022
Genetesis
Machine Learning Engineer II
August 1, 2018 – August 1, 2020
Genetesis
Machine Learning Engineer
November 1, 2016 – August 1, 2020
Gofind.ai
Machine Learning Intern
August 1, 2016 – November 1, 2016
Rochester Institute of Technology
Student
August 1, 2014 – May 1, 2016
L1-Grassman Manifolds for Object Tracking (MATLAB)
December 1, 2015 – Present
•Used different sequences from the Visual Object Tracking (VOT) challenge 2014 dataset for tracking. •Implemented a robust tracker using L1-Grassman manifolds where each class subspace is mapped onto the manifold using L1-PCA algorithm and projection kernels projects the manifold onto Hilbert space. •After reducing the dimensionality, classification algorithm gives the final tracking result.
Credit Risk Analysis and Prediction System (Python/R)
November 1, 2015 – Present
- Analyze the German credit data set and classify customers as good or bad credit risks using a set of categorical/symbolic attributes. - Applied and compared different classification algorithms such as Logistic regression, Gradient boosting, Nearest Neighbor, Random forests, Linear/Quadratic Discriminant analysis and an Ensemble classifier. - Highest accuracy of 78% obtained using Random forest classifier after performing feature selection and data analysis.
Object Recognition using HOG / SVM vs. CNN (NVIDIA DIGITS)
November 1, 2015 – Present
It involves comparison of the traditional HOG/SVM approach with the state of the art Deep learning approach using the GoogLeNet Inception architecture on recognition of 102 different objects using the Caltech 101 dataset. Used the NVIDIA DIGITS GPU training system with Cafee as the deep learning framework for the recognition task using the GoogLeNet incarcination of the inception architecture of Google as the network model. The validation accuracy of around 88% using the deep learning approach was much better than the 50% accuracy for the traditional HOG/SVM approach.
CRAS: Customer Review Analysis System( R/ Python)
September 1, 2015 – Present
Analyze the customer review sentiments and classify them as positive or negative. Classification based on Logistic regression and compared with 4 other classification algorithms namely SVM, KNN,Random forests and Naive Bayes for performance and accuracy, highest being 85%. Amazon data-set used for the review sentiment analysis. Based on the sentiment analysis, the system helps the company in analyzing the products which need improvement and it also informs about the products well received by the customers.
Face Recognition using Support Vector Machines (MATLAB)
May 1, 2015 – Present
In this project, eigenfaces which is a part of principal component analysis (PCA) are used as features. The advantage of using PCA is that is reduces the dimensionality of the features and requires less amount of processing time. For the classification purpose we have made use of Support Vector Machines along with Polylinear Kernel. We have compared the performances of SVM on the ORL database of faces and the Yale Data set. ->>An accuracy of 82.5% in case of the ORL database and an accuracy of 88.89% in case of the Yale database of faces was obtained. ->>We used a different technique called as modular PCA and the accuracy obtained was more than the one obtained by using PCA. An accuracy of 84.68 % in case of the ORL database and an accuracy of 93.89 % in case of the Yale database of faces was obtained by using Modular PCA. ->>For using the Support Vector Machine we have used the libsvm library.
Color Segmentation using kmeans clustering using MATLAB
April 1, 2015 – Present
1. Performed color segmentation of the color image based on kmeans clustering of its intensity component where each segment in the segmented image has a different gray tone where the value of k is selected based on the number of important regions in the image. Also, the value of k is varied to observe the effects of choice of k on segmentation results. 2.Performed segmentation of the color image based on kmeans color clustering where each color pixel is in RGB space. Here, each segment in the segmented image has a different color. And, the value of k is also varied in the same manner as above to observe the effects of choice of k on results of segmentation. 3. Observed the difference between Segmentation of Color vs Grayscale image based on kmeans clustering.
Image Retrieval using color histogram features (MATLAB)
April 1, 2015 – Present
1. A 192 bin quantized color histogram of each image in the given color image dataset is computed and stored. 2.A test or a model image is selected from each class of the database and 5 closest or best matches are computed based on the Histogram Intersection distance method.
Image Retrieval using Texture features (MATLAB)
April 1, 2015 – Present
1. In this, from the given texture image dataset, the intensity (grayscale) component of each image is used to compute a 9 dimensional texture feature vector based on the Law filter masks. Before filtering, illumination compensation is done to adjust the contrast and intensity of each texture image using Histogram equalization or stretching. 2. A model texture image is selected from each class and 5 best or closest matches are obtained based on the Chi Squared distance of the feature vectors.
Kmeans Clustering
November 1, 2014 – Present
CPU and GPU implementations for Kmeans Clustering were executed using the provided header file on a set of Datapoints, producing a result in which a group of datapoints are assigned to any one of the 3 clusters. GPU implementations were analyzed and compared with the CPU implementations for parameters like Computation Time and cluster assignments. Constant memory was used for the clusters in GPU implementation. The CPU was still involved during the execution of the program to coordinate the number of iterations that were required in the GPU implementation. The results showed that the GPU is much faster than the CPU implementation as the size of datapoints increase.
Parallel Implementation of a Hodgkin-Huxley Neuron Model using MPI
November 1, 2014 – Present
• Given the sequential implementation of the neuron model, a parallel version of the same was implemented using Message Passing Interface with basic communication between master and various slave processes. • Effect of Dendrite length and the number of dendrites on Execution times were recorded. • Studied the effects of parallel implementation and load balancing on Computation time and Parallel vs. Sequential Speedup
JPEG Compression in CUDA
November 1, 2014 – Present
• CPU implementation and GPU implementation of JPEG Compression of an image was compared for performance measuring Speedup of GPU vs. CPU. • Used CUDA C to write different kernels for the different steps involved in JPEG Compression for the parallel GPU implementation. Achieved very low Compression Ratio by dividing the image into Minimum coded units and then applying the algorithm on each unit.
Parallel Implementation of a Ray Tracer
October 1, 2014 – Present
• With the help of Message Passing Interface(MPI) with basic sends and receives for communication between master and various slave processes, a parallel code of a Ray Tracer was implemented using different Static partitioning schemes and Dynamic partitioning scheme with a Centralized Task Queue. • Effect of task grain size and job allocation methods were observed on the performance of parallel program comparing execution times, C-to- C ratio & Speedups of all the different partitioning schemes
Gauss Jordan Elimination
October 1, 2014 – Present
CPU and GPU implementations for Gauss-Jordan Elimination were executed using the provided header file on a matrix, producing a result that is in reduced row echelon form. CUDA implementations were analyzed and compared with the CPU implementations for parameters like Computation Time, Speedup and Relative error. After this, an additional procedure of Partial Pivoting for the CPU was done to enhance precision and then was compared with the original CPU implementation without Partial Pivoting method for Speedup, Error and Computation time. The results showed that the implementation is not embarrassingly parallel because each iteration is dependent on the previous iterations and so it is very complex to deal with the GPU implementation due to the data dependencies.
Median Filtering
October 1, 2014 – Present
CPU and GPU implementations for Median Filtering were executed using the provided header file on an input image , loading the image (using the provided Bitmap class) and perform median filtering first with the CPU, with the GPU implementation using global memory, and finally an implementation using shared memory, comparing the results and performance using the timing techniques. The results showed that the GPU implementation using shared memory takes less time as compared to the GPU implementation using global memory and also both the GPU implementations are much faster than the CPU implementations for a 3 x 3 Median Filtering.
ECG ACQUISITION AND COMPRESSION TECHNIQUE
August 1, 2013 – April 1, 2014
-An ECG acquisition kit was developed -ECG signal of the patient was obtained and processed it using MSP430 Launch pad to get the digital values of the signal. -The digital signal data that was obtained was converted using MATLAB to ECG signal data file -This data file was used to perform various compression techniques like discrete sin transform, discrete cosine transform, fast-Fourier transform, discrete wavelet transform and the results was obtained -Quality analysis was performed for the different compression techniques and the result were compared based on different parameters -Parameters that were considered were compression ratio, percent root difference,signal to noise ratio,maximum amplitude of error -Decomposition of 12-lead ECG signal was also performed.
Cultural Fit Analysis
The candidate's project diversity, ranging from credit risk analysis to medical imaging and visual search, indicates adaptability and a broad interest in applying ML across different domains. Their progression through various ML Engineer and Research Scientist roles in different companies (Genetesis, Arya.ai, Lumen5, Workday) suggests a proactive approach to career growth and continuous learning, aligning with a dynamic work environment. The focus on real-world problem-solving and performance optimization in projects also aligns with a results-oriented culture.
Soft Skills & Operational Fit
The candidate's experience descriptions highlight leadership, collaboration with product and sales teams, and a focus on delivering business value, indicating strong operational fit and communication skills. The project descriptions demonstrate a methodical approach to problem-solving and performance comparison.