Shailesh Nanisetty

Hi, I'm Shailesh Nanisetty.

A

A self-driven, resilient, and passionate learner with a curious mind thriving on solving challenging real-world problems through machine learning and generative AI.

About

I am a Master of Science in Applied Computing (Artificial Intelligence) student at the University of Toronto, with a strong background in Industrial Engineering from IIT Kharagpur. I enjoy problem-solving and always strive to bring 100% to the work I do. My main focus is on developing cutting-edge ML driven solutions through applied research. With diverse experience in professional internships and research projects in areas pertaining to Generative Machine Learning, Large Language Models (LLMs) and Computer Vision, I have honed my skills in Python, PyTorch, TensorFlow and various ML frameworks utilizing it to solve interesting problems. I am particularly passionate about developing complex applications that solve real-world problems impacting millions of users. Please reach out to discuss, meet or collaborate 😃

Languages: Python, JavaScript, C++, R, Matlab
Databases: MySQL, PostgreSQL, MongoDB
Libraries: OpenCV, NLTK, Hugging Face,Numpy,Pandas, Statsmodels, Scipy, Scikit-Learn, PySpark
Frameworks: Flask, Django, Node.js, Keras, TensorFlow, PyTorch, Bootstrap
Tools & Technologies: Git, Docker, AWS, Heroku, JIRA

Actively seeking Full-time opportunities to work in a challenging position combining my skills in Machine Learning, which provides professional development, interesting experiences and personal growth.

Experience

Servicenow Research

Visiting Researcher

Developed AgentAda, an innovative LLM-powered analytics agent that automates method selection for specialized insights.
Architected a 3-part system integrating a question generator, RAG-based skill matcher & code generator to streamline data analysis.
Created KAGGLEBENCH, a benchmarking tool of expert-curated notebooks to validate AgentAda’s analytical performance.
Secured 48.78% human approval demonstrating AgentAda’s superiority over traditional methods with an LLM-as-a-judge strategy.
Developing BigInsightBench - The 1st Global Benchmark for Multimodal Analytics to enhance LLM-driven decision making.
Tools: Python, LLMs, Prompt Engineering, Git

February 2025 - Present | Toronto, Canada

Synapsis Medical Technologies Inc.

ML Developer Intern

Developed a React app integrating 3 ML models to get real-time facial expressions, head poses and heart rates with a 75 ms latency.
Designed SSL model leveraging contrastive learning with triplet loss to improve representation learning for the aforementioned models.
Implemented hard negative mining in triplet loss to enhance embedding separability, reducing facial emotion classification loss by 15%.
Pioneered a multimodal monitoring system to analyze interaction of emotional, physical & physiological states for tailored insights.
Tools: Python, PyTorch, Tensorflow, React, Docker

Aug 2024 - Feb 2025 | Toronto, Canada

Balyasny Asset Management

Data Science Associate

Developed an Apache Airflow DAG for BAM Elevate, automating data population processes & improving workflow efficiency by 30%
Structured multi-level datasets from investor & deal interaction feedback, enhancing analysis speed by 25%
Optimized SQL queries for faster geographical expansion insights, reducing extraction & aggregation time by 40%.
Developed dynamic Streamlit dashboards to analyze investor feedback & guide decisions on geographical expansions in real-time.
Tools: Python, PostgreSQL, Streamlit, PySpark, Databricks

May 2024 - Aug 2024 | Toronto, Canada

16 Bit

Generative ML Intern

Developed PaddleOCR-ViT model to extract patient details from DXA sheets, achieving 92 % accuracy on test set of 1200 samples
Performed zero-shot & few-shot prompting on Llama2 7B & Zephyr 7B, achieving ROUGE-L score of 45.2 in report summarization.
Fine-tuned models using PeFT-LoRA techniques enhancing contextual accuracy of report summaries with a ROUGE-L score of 48.3.
Tools: Python, PyTorch, HuggingFace, Transformers, Git

November 2023 - February 2024 | Toronto, Canada

Bharat Smart Services

Computer Vision Intern

Developed & deployed custom CNN models, achieving 91% accuracy on test set, to classify electric meters & extract readings using YOLOv5.
Delivered a solution for a POC involving an Android app for real-time electric meter reading & automated billing, streamlining the process for end-users.
Tools: Python, Tensorflow, Android Studio, Docker, Git

May 2021 - Aug 2021 | Hyderabad, India

Projects

                
Visual Question Answering
                  An attention-based classification model that aims at generating an answer for a given input image.
                
AccomplishmentsIncorporated Convolution Neural Networks (CNN) for extracting image features and Long Short Term
                    Memory for extracting question
                    embeddings.
Tested the model on the COCO dataset, abstract scenes images, and got 81% overall accuracy on the
                    VQA evaluation metric.

Tiny-HR
                  Tiny-HR: An interpretable machine learning pipeline for heart rate on low power edge devices
                
AccomplishmentsBuilt a hybrid FFNN-CNN pipeline that extracts heart rate from pressure data acquired on low-power ESP32 device.
Wrote C++ Scripts for suitable deployment of implemented PyTorch models onto ESP32 edge device.
Proposed method cuts energy and time inference by 82 % & 28 % compared to state of the art methods.

Few-Shot Segmentor
                  An Improvised sequential few-shot segmentation through texture bias removal
                
AccomplishmentsDesigned a Few-Shot CNN algorithm for segmenting low-labelled images by reducing perceptual bias.
Incorporated set of Difference of Gaussians and bi-directional ConvLSTM algorithm in the framework.
Performance measured in mean IOU shot up by 6.26 % & 1.2 % for 1 & 5 shot cases respectively.

Bank Lending Optimization
                  A hybrid model of Genetic & Simulated-Annealing Algorithms to optimize bank objectives on the loan portfolio.
                
AccomplishmentsBuilt a hybrid model of Genetic and Simulated-Annealing Algorithms to optimize bank objectives on the loan portfolio. 
Brainstormed on maximizing bank profit & minimizing probability of bank default in search for dynamic lending decision.
Obtained fitness function value and an array of binary digits (1-Customer selected and 0-not selected) as the final solution. 
Achieved a best optimal fitness value of 3.1854 with hybrid of genetic and simulated annealing algorithm compared to Genetic Algorithm’s 2.723.

Electric meter reading App
                  An optimized Flask based App for Electric Meter reading.
                
AccomplishmentsBuilt a custom deep learning model to classify two sets of meter & non-meter images using CNN in Tensorflow (GPU). Employed Dropouts and L2 Regularization for better performance on the test set and avoid overfitting over training data. Used Minibatch Gradient Descent for better training and achieved 77.2% Training and 83.1% Test Accuracy, 
                    which improved to 93.1% with InceptionV3 Model. Finally, Created & Deployed a flask-based App that does Electric Meter Classification and further gives reading through image recognition.

Predictive Process Mining
                  Built a hybrid framework of clustering and supervised algorithm for classifying the process data of patients in a hospital.
                
AccomplishmentsCreated a tool for automatically labelling process data of patients in a hospital using a set of Linear Temporal Logic rules; Applied Gaussian Mixture Model & DBSCAN resp. for clustering data into similar groups after proper data pre-processing.
                    Employed Decision Tree & Random Forest (with each of the above clustering techniques) to classify the processes into Normal and Deviant categories; Achieved best results with GMM & Decision Tree Combination.