New England Northern AI on Edge: Novel Post-Training Quantization for Education Applications

Overview

Could a personalized, portable AI tutor transform education and improve outcomes for students in disadvantaged communities? Advancements in open -source large language models (LLMs), particularly multimodal models that can understand images and text, enable AI-driven learning by giving students personalized feedback while addressing privacy and sustainability concerns. However, running these models on consumer edge devices like a cell phone remains cost - prohibitive. I hypothesize that AI can be made more eff icient for use on a phone through an improved algorithm that decomposes connection strengths of a neural network into two parts, one within a range that can be exploited to reduce memory footprint. This would allow me to fit a powerful AI model onto a phone while retaining high accuracy. My novel post-training quantization method outperforms other quantization techniques, achieving up to 3.27x higher speed, superior accuracy, and at least the same level of size reduction. It compresses a cutting -edge 8-billion- parameter model from 16 GB to 8.16 GB RAM, reducing its size by 49%. Integrated into my custom open -source inference engine written in Rust called mistral.rs, this approach powers Edge(u)cation, an AI tutor app I created for mobile devices. To validate its impact, I then deployed Edge(u)cation in several sample settings including math and engineering bridge structure analysis experiments through real -time, AI -driven feedback. In conclusion, this work demonstrates a scalable, cost-effective solution for personalized learning, fostering STEM engagement in under- resourced communities, enabling usage offline or off -grid. I published all codes and models in open-source for anyone to use and build on: https://github.com/EricLBuehler/edge-u-cation. Modulation of TLR2-Mediated Innate Immunity Mitigates Inflammatory Pathology in Alzheimer’s Disease Ethan Liu Phillips Exeter Academy, Exeter, NH Alzheimer’s disease (AD) is an irreversible neurodegenerative condition marked by the progressive loss of memory and cognitive function. The risk for AD increases exponentially with age. This study aims to explore the molecular mechanisms that differentiat e healthy aging and AD-associated aging. In the experiments, I combined single -cell transcriptomic analyses of multiple AD patient datasets with drug validation using the following AD experimental models: transgenic strains of the nematode worm Caenorhabditis elegans (GMC101 strain expressing amyloid-β protein in body wall cells and N2 WT control strain) and mouse cortical neuron cultures. From bioinformatic analyses, I discovered several differentially expressed genes (e.g. TLR2, HSD11B2, CXCR4) that potentially drive the pathogenesis of AD. I examined these genes through several in vivo and in vitro experimental methods. Immunohistochemical staining and brightfield imaging of neurons treated with C29, a potent TLR2 antagonist, revealed increased cell survival and more neural processes compared to controls. Furthermore, C. elegans lifespan and movement assays demonstrated that C29 significantly boosted the lifespan and motility of AD worm models. Through comprehensive bioinformatic analysis and experimentation, I conclude that TLR2 antagonist C29 may be a novel disease -modifying drug candidate for AD. This stu dy provides valuable insights into the pathogenic nature of upregulated TLR2 in AD. These conclusions will support future research and develo pments of disease -modifying drugs for AD, promote healthy aging, and improve healthspan and lifespan.

Competition history

• JSHS 2025 Category not listed

Resources

• National JSHS abstract book ↗

Related projects

Closest projects by meaning, across every fair and year in the corpus.