My Start-up Neuroflux: Building Global Innovation in AI and Neuroscience
The journey began at the New York Academy of Sciences where I was so fortunate to meet my colleagues and co-founders of what has now become Neuroflux. It all started with us wanting to solve the problems of ethical AI and explainability. What began as a spark of curiosity during an international ethical AI challenge grew into a powerful global collaboration. In September 2024, I co-founded Neuroflux (https://www.neurofluxai.org/), a startup that bridges continents and disciplines to tackle a tough challenge: how should AI be used to more accurately read a brain scan, highlight more accurately the presence of Glioblastoma Multiforme (GBM), an aggressive form of brain cancer, and provide an explanation.
Together, we developed a multimodal Grad-CAM-enhanced U-Net deep learning model designed to improve tumor segmentation accuracy while making AI’s decision process transparent and explainable.
Our team is spread across time zones within the US and across the world. Despite our geographic distance, our collaboration has thrived through constant communication, shared purpose, and mutual respect. We learned to write code, test models, and publish papers across borders with each late-night call and early-morning meeting reminding us that innovation has no boundaries.
We are very proud and at the same time humbled that our work has been widely recognized. We were awarded a publication scholarship by the International Journal of High School Research, where our study appeared in the June 2025 issue. We presented at Harvard OpenBio’s “Revolutionizing Life: Frontiers of Biotechnology” conference, and our project was accepted at the 7th International Conference on Neuroscience, Brain Disorders and Therapeutics.
The entrepreneurial side of our journey took off as well. We placed 2nd globally at the 2025 Wix Creators of Tomorrow Challenge (earning a $2,000 award) and being selected for the prestigious Wix Semifinals Cohort. We were also ranked 36th among 2,000+ global teams from 160+ countries in the Moonshot Pirates “Shape the Future” Challenge, qualified for the UN SDG Challenge public voting stage, and won first place at Harvard College Vision’s Global Health & Leadership Pitch Competition, outperforming 86 other international finalists. We are most excited to be one of three finalists at the Future Port Youth conference. The awards will be presented on the 13th of November in Prague and we are very excited to be there in person!
Our goal with Neuroflux was to ensure that the advances we have made are widely available. We are pleased to say that our library has had 7000+ downloads. For me, Neuroflux has been more than a research project. It is a startup that is based on sound innovation. Its given me the opportunity to work with friends who are innovators across the world and together, make an impact.
Certificates: https://drive.google.com/drive/folders/1yIffXjFLKFiXPizMi51hVlHBOnh-z_Ni
My IJHSR Publication - Neuroflux
Neuroflux: An Interpretable U-Net Model for Glioblastoma Multiforme Segmentation Through Grad-CAM
DOI: 10.36838/v7i6.64
ABSTRACT: Glioblastoma multiforme (GBM) is one of the most aggressive malignant brain tumors, with early intervention through diagnosis and treatment being critical to improving the low median overall survival rate of 15 months. Manual anatomical segmentation is problematic because it is prone to inconsistencies from interobserver variation. Current models are highly accurate but lack computational efficiency and multimodality, which limits the clinical applicability of such segmentation tools in hospital environments. Neuroflux is a multimodal U-Net-based deep learning model for GBM segmentation using the BraTS 2020 dataset. Adapted from a baseline model, we enhanced the preprocessing of raw data through MinMax scaling and applied our custom hybrid loss strategy to maintain high levels of accuracy while minimizing false negatives. With further improvements in regularization techniques, Neuroflux was also fine-tuned to incorporate targeted modifications and optimized dropout rates,which improved the training strategy. Overall, an accuracy of 99.01%, precision of 99.11%, sensitivity of 98.85%, and specificity of 99.7% was achieved while significantly improving clinician interpretability. Our high-precision and efficient modified architecture highlights the potential for new lightweight and interpretable models to enhance diagnostic reliability and treatment planning.