Mahadev Sunil Kumar

Introduction

Hey, I'm Mahadev! I just finished my undergrad in Artificial Intelligence from Amrita University in 2025. I'm highly passionate about doing research in the field of optimization of AI algorithms, and I'm currently working on a project in that regard. I'm also interested in Neuromorphic Systems and I'm currently searching for PhD opportunities in that field. I would love to collaborate with like-minded individuals who share the same passion for research as me and dedicate their lives to research.

Research Interests

• Neuromorphic Systems
• Optimization of Algorithms
• Computer Vision
• Edge Computing

Education

• B.Tech in Artificial Intelligence, Amrita University (2021-2025)
• Indian School Certificate, Loyola School (2019-2021)
• Indian Certificate of Secondary Education, Loyola School (2008-2019)

Experience

• Advanced App Engineering Analyst (October 2025-Present), Accenture, Bengaluru, India
• Teaching Assistant (May 2025-June 2025), CIR, Amrita University, Amritapuri Campus
  • Delivered sessions in competitive programming, algorithms, and efficient coding to the B.Tech. 2022–2026 batch; emphasis on complexity analysis, data structures, and problem patterns.
  • Designed and taught Code Hour, a placement-oriented module covering greedy, DP, graph algorithms, recursion/backtracking, bit manipulation, and code hygiene.
  • Mentored 143 students; created graded practice sets, solution walkthroughs, and rubric-based feedback to improve correctness and speed under time constraints.
• Undergraduate Student Researcher (September 2023 - June 2025), CIR, Amrita University, Amritapuri Campus [Link to Project Details]
• Project Intern (August 2023-October 2023), ISRO Inertial Systems Unit, Trivandrum, Kerala, India [Link to Project Details]
• Intern (June 2023-October 2023), ACM Student Chapter, Amritapuri Campus
• Intern (August 2022-September 2022), Kerala State Film Development Corporation, Trivandrum, Kerala, India

Projects

SlimEdge: Lightweight Distributed DNN Deployment on Constrained Hardware (September 2023-Present)

Preprint arXiv Link: Link to be added
Supervised by: Dr. Amitava Mukherjee (Birla Institute of Technology, Dubai), Dr. Arnab Raha (Intel Corp.), Dr. Debayan Das (Indian Institute of Science), Dr. Gopakumar G (Amrita University)


• Designed dynamic filter-pruning for CNNs under memory/latency constraints; extended to Multi-View CNNs with per-view importance estimation using XGBoost.
• Built reward functions combining accuracy preservation, model size, inference time, device memory caps, and per-view salience.
• Implemented single-view optimization with a Genetic Algorithm; scaled to all views using NSGA-II for multiobjective trade-offs across accuracy, size, and latency.
• Developed sampling strategies (importance- and device-aware) using Beta distribution parameterizations for pruning vectors.
• Used root-finding (Newton-Raphson, Brentq, TOMS748) to solve budget-constraint equations and align pruning targets with device memory ceilings.
• Deployed and profiled on heterogeneous nodes (Raspberry Pi, NVIDIA Jetson Nano, Apple Silicon).
• Implemented device-aware partitioning heuristics for distributed inference; validated robustness via cross-device accuracy and latency variance analysis.

YOLOv7 based Gauge Reading System (August 2023-October 2023)

Supervised by: Ouseph P. (IISU, ISRO), Durairaj R. (IISU, ISRO)

• Built a YOLOv7-based system for humanoid analogue-gauge reading; curated and annotated a custom dataset with variable lighting, glare, and dial geometries.
• Engineered post-processing for pointer-angle regression and scale normalization; achieved ~98% accuracy on held-out gauges; created an inference workflow for deployment.
• Deployed trained model on an NVIDIA Jetson as well as optimized the model using ONNX and TensorRT.

• Speech Emotion Detection: (May 2024-June 2024)
  • Built MFCC and spectrogram pipelines; trained LSTM classifier for multi-class emotions; performed augmentation and class-imbalance handling.
  • Evaluated with stratified splits; reported precision/recall/F1 per class and confusion matrices.

• Reinforced Model for Chess: (February 2024-July 2024)
  • Implemented self-play DQN with episodic reward shaping and n-step returns; stabilized training with target networks and experience replay.
  • Benchmarked against heuristic baselines; analyzed policy improvement curves and move-quality statistics.

• Improving SDN Security against DDoS attacks: (December 2023-February 2024)
  • Engineered flow-level features; trained SVM to detect DDoS in SDN settings; validated on held-out traffic traces.

Publications

Preprints

• SlimEdge: Lightweight Distributed DNN Deployment on Constrained Hardware: Link to be added

Skills

• Programming Languages: Python, Java, C++
• Machine Learning: TensorFlow, PyTorch, Scikit-Learn
• Audio Engineering: Logic Pro, Studio One
• Video and Graphic Design: DaVinci Resolve, Final Cut Pro, Adobe Suite
• Soft Skills: Teamwork, Communication, Problem-Solving