|

osdev

12-01-2024 -- Present

Technologies: x86 Assembly

    • WIP
    • Developed a minimal operating system from scratch using Assembly, and C.
    • Focused on low-level system programming and hardware interaction.

veri-nn

10-01-2024 -- 11-01-2024

Technologies: Verilog, ModelSim, PyTorch

    • Engineered a high-performance neural network library in Verilog with modular, scalable components, meeting 4MB on-chip memory constraints for deploying ML architectures on a DE1-SoC FPGA, achieving 2.27ms inference speed.
    • Trained a neural network in PyTorch, utilizing quantization techniques to optimize memory usage with int32 precision, achieving 80% validation accuracy on the MNIST dataset.

hydrogen-compiler

09-01-2024 -- Present

Technologies: C, C++, x86 Assembly

    • WIP
    • Developed a compiler for a custom programming language.
    • Implemented lexical analysis, parsing, and code generation.

phishing-reel

08-01-2024 -- 08-31-2024

Technologies: Python, Jupyter Notebook

    • NewHacks 2024
    • Developed a phishing email detection system using transformers.
    • Analyzed email content and metadata to identify phishing attempts.

flow-sense

08-01-2024 -- 08-31-2024

Technologies: React, Django, PostgreSQL, GPT-3.5

    • HackThe6ix 2024
    • Developed a reading tool that enhances productivity by integrating LLM API-based word definitions and inline note-taking, reducing research time by 40%.
    • Successfully preserved the original report formatting in PDF processing, maintaining 95% accuracy in layout retention.
    • Built and deployed within a 36-hour period during Hack the 6ix 2024, achieving a functional MVP with time to spare for feature enhancements.

deepfake-audio-detection

07-01-2024 -- Present

Technologies: Python, Jupyter Notebook, PyTorch

    • WIP
    • Developed a deepfake audio detection model using machine learning techniques.
    • Analyzed audio features and trained models to distinguish between real and fake audio.

robot-line-challenge

06-01-2024 -- 07-01-2024

Technologies: C++

    • Implemented a robot line-following algorithm in C++.
    • Focused on sensor data processing and motor control.

mini-torch

05-01-2024 -- Present

Technologies: C++, CUDA, LLVM

    • WIP
    • Recreated PyTorch’s object-oriented design with C++ and CUDA, implementing core functionalities such as LinearLayer, ReLU, Softmax, and CrossEntropyLoss, enabling an intuitive setup for deep learning training loops similar to higher-level languages.
    • Optimized CUDA kernel functions for matrix multiplication through memory coalescing and tiling, and further utilizing kernel fusion with ReLU, to achieve a 5x reduction in training time.

image-captioning

05-01-2024 -- Present

Technologies: Python, PyTorch

    • WIP
    • Engineered a deep learning model combining Computer Vision and NLP, using PyTorch to integrate CNNs with Vision Transformers/LSTMs, achieving a max BLEU-4 caption score of 11.0.
    • Designed models using transfer learning with InceptionV3 and VGG16, improving BLEU caption scores by 0.2.

dish (dope interactive shell)

05-01-2024 -- Present

Technologies: C

    • WIP
    • Built a shell program in C with command parsing and execution capabilities.
    • Implemented tokenizer function and lexing function

trash-talk

01-01-2024 -- 01-31-2024

Technologies: Arduino, Flask, HTML, CSS, JavaScript

    • UTRAHacks 2024
    • Developed an Arduino-based waste sorting system that accurately categorizes trash into garbage or compostable materials, achieving a sorting accuracy of approximately 85%.
    • Integrated a web interface with real-time data visualization, reducing manual sorting time by 50% for users.
    • Implemented hardware components such as an ultrasonic sensor and moisture sensor, optimizing the sorting mechanism to handle up to 30 items per minute.

deep-learning-from-scratch

12-01-2023 -- Present

Technologies: C/C++, CUDA, Python, PyTorch, NumPy

    • Spearheaded Neural Networks, CNNs, RNNs, LSTMs, and Transformers from scratch using OOP principles, increasing code implementation efficiency by 50%.
    • Programmed implementations to utilize CUDA acceleration using NVIDIA’s C++ libraries, resulting in a 5x speedup in training.
    • Produced educational resources with practical examples, instructing over 5+ individuals on deep learning theory.

ai-art-detection

12-01-2023 -- 06-01-2024

Technologies: Python, PyTorch, NumPy

    • Co-developed a CNN-based deep learning model for AI-art detection using PyTorch, achieving a maximum 90% test accuracy.
    • Leveraged transfer learning with pre-trained models such as VGG16 and ResNet50, reducing model loss by 6%.
    • Evaluated 10+ architectures proposed in research papers to identify the best-performing model.
You've reached the end of the page.