Shantanu Welling

CacheBoost: Memory Hierarchy Optimizations

Recognized as the BEST PROJECT for exceptional insights, presentation, code and content delivery among 60+ teams as a part of undergrad course CS230: Digital Logic Design & Computer Architecture.

• Optimized Best Offset Data Prefetcher by removing delay queue, increasing prefetch distance, prefetch degree and the number of candidate offsets tested per memory access using ChampSim microarchitecture simulator.
• Integrated Hawkeye & Mockingjay non-heuristic cache replacement policies and tested optimizations in their penalty factors, eviction priority of blocks and treatment of prefetch/demand accesses & cache averse lines.
• Experimented with cache sizes & set-way associativity for inclusive, exclusive & non-inclusive cache hierarchies.

Medical Image Refinement

• Denoised images with Markov Random Field prior through Iterative Conditional Mode by using a discontinuity-adaptive Huber potential function.
• Measured similarity between shapes by aligning equicardinal pointsets through similarity transform using Procrustes distance to determine the optimal rotation in the pre-shape space & modes of shape variation.
• Reconstructed CT images by Algebraic Reconstruction Technique and performed Radon Transform & filtered back-projection using Central Slice Theorem with Shepp-Logan & Ram-Lak filter on X-Ray phantom images.

Hypervisors, Containers & Linux Kernel

CS695: Virtualization & Cloud Computing

• Implemented a hypervisor with several hypercalls, hypercall handlers & a VM scheduler using Linux KVM API.
• Developed a container management tool with network configuration features utilizing namespaces & cgroups.
• Created a virtual device driver & modified the linux signal handler through Linux Kernel Modules & IOCTL.
• Estimated the working set size of a process by monitoring the idle & accessed page flags via kernel page bitmap to measure the Active Physical Memory footprint.

Rail Planner

CS293: Data Structures & Algorithms

• Developed a one-stop application for Rail Travel which includes booking a journey, viewing and managing current journeys, maintaining reviews for all journeys.
• Integrated data structures (in C++) for optimizing app processes, utilizing Linked Lists, Dictionaries for storing journeys, Binary Search Trees for price comparisons, Tries for storing station names, Priority Queues for filtering journeys & reviews according to ratings & Knuth-Morris-Pratt algorithm for querying reviews by keywords & stations.
• Implemented BFS, DFS, Kruskal's & Dijkstra's graph algorithms for finding journey routes between stations.

Relation Networks

We explored the use of Relation Networks in the context of Visual Question Answering. We implemented the Relation Network architecture for relational reasoning proposed by Santoro et al. and evaluated its performance on the CLEVR dataset as well as a variation of the CLEVR dataset called Sort-of-CLEVR.

We try to replicate the results obtained for models corresponding to three different input specifications, namely the input with images and encoded questions, the input with and unencoded questions, and the input with object descriptions and unencoded questions. Thus, demonstrated the model’s flexibility to treat unstructured inputs as a set of objects for relational questions.

SuperPixel Graph Cuts

• Classified image regions into background/ foreground using network flow graph model over superpixel nodes.
• Utilized CIELab colorspace and Simple Linear Iterative Clustering algorithm (SLIC) for SuperPixel generation.
• Designed a probabilistic edge weight model to execute Boykov Kolmogorov algorithm for min s-t cut partition.
• Integrated interactive GUI to take user markings of image regions that serves as a prior for the edge weight model.

FastChat

• Connected a network of clients using socket programming by designing a communication software in Python.
• Utilized PyCrypto for E2E encryption of messages for user security and privacy protection & MySQL database to store SHA256 hashed user passwords.
• Integrated group chat, undelivered messages for offline clients, load balancing server & image transfer features.
• Experimented on high throughput of servers & low latency of message deliveries through network traffic analysis.

Image Generation & Variation Analysis

Analyzed the principal mode of variation for each digit from the MNIST Database of 28 x 28 images of handwritten digits and performed PCA & LDA to optimally reduce image dimensionality to 84 and reconstruct the image.

Developed a Python program that takes in images of fruits, utilizes Principal Component Analysis and hyperplane fitting to generate random sample representative images of fruits absent in the training set.

Stock Market Trend Predictor using Deep Learning

Forecast stock price trends by implementing a CNN-LSTM model using Momentum & Paired Switching financial indicators in Python (TensorFlow) as a part of Seasons of Code 2022 conducted by the Web & Coding Club, IIT Bombay.

Oil Price Analysis

Inspected 3 oil companies’ stock prices & WTI oil barrel prices via t-test & regression analysis to estimate trends as a part of our course project for AE102: Data Analysis & Interpretation.
Performed hypothesis testing & confidence interval estimation, computed correlation matrix to better understand the data.

Brain Tumor Detection

Implemented brain tumour segmentation & detection model to perform diagnosis & estimate survival rate from multi-modal 3D MRI images using 3D U-Net architecture.

Sustainable Tourism

Worked on a Sustainable Tourism project and case study for the final level of Dr. Homi Bhabha Junior Scientist Science Talent Search Competition 2017-18 for which I was awarded the Silver Medal at the State level by Dr. Anil Kakodkar (an Indian nuclear physicist and mechanical engineer. He was the chairman of the Atomic Energy Commission of India, he was the Director of the Bhabha Atomic Research Centre as well).