Melvin Orichi Socana

I am a graduate from Stanford University (Class of 2024) with a Master's degree in Electrical Engineering, concentrating on hardware, software, and machine learning systems. My advisor was John M. Pauly at Stanford University. I obtained my Bachelor's degree from North Carolina State University (2018) in Computer Engineering & Electrical Engineering.

In addition to my core studies, I have taken substantial coursework in machine learning, making me adept at designing and implementing hardware systems, software systems, and machine learning models, as well as handling the distribution and concurrency challenges these systems require.

My professional experience includes internships at Amazon as a Software Development Engineer (SDE) and at Johnson & Johnson as an Electrical Engineering intern on the R&D team. I have also engaged in research experience during both my undergraduate and graduate studies. I am passionate about creating efficient, scalable, and innovative solutions, and I am currently seeking a role where I can leverage my skills and contribute meaningfully to impactful projects.

Email  /  Resume  /  Github  /  LinkedIn

Projects

A few-shot sequence-to-sequence semantic parser for Wikidata. We introduce a preprocessing step that utilizes the generative capabilities of GPT to provide enriched contextual descriptions for entities.

The paper describes the development and improvement of a CNN-based model for predicting human movement states using accelerometer and gyroscope data. An ensembling approach combining predictions from 7 models, a custom temporal filter to stabilize predictions, and hyperparameter optimization via random search, achieving a 92.4% accuracy with reduced noise and prediction spikes is used. We obtain 3/29th place for this competition.

Built a robust image classifier using ResNet50 and transfer learning to identify artists from artwork, achieving an F1 score of 0.835 and placing first in a Kaggle competition.

Designed and developed a custom LED printed circuit board from PCB creation in KiCAD to writing integrated control software.

We develop a Sentence-BERT model that achieves performance on sentiment analysis, paraphrase detection, and semantic textual similarity tasks.

A static wear-leveling technique for memristor crossbar array architectures to enhance endurance by evenly distributing write/erase cycles across the array, aiming to mitigate wear imbalance and improve the reliability and lifespan of memristor-based memory systems.

RINN-UOD leverages randomly initialized neural networks for unsupervised object detection by using CNN-based feature extraction, clustering for region proposal, and non-max suppression to produce class-agnostic bounding boxes without supervised or pre-trained models.

We develop our own blockchain via an independent implementation of a node in TypeScript.

Coursework

Website based on Joe Barron's website.