Agentic AI & GenAILightning talk20min
Beyond the Cloud: Architecting Zero-Latency Edge AI Pipelines for Cognitive Accessibility
This proposal presents a privacy-first, offline assistive AI system for education that avoids cloud latency and data risks. Using Mount AI Scholar, it covers edge-based small language models, optimized phoneme-grapheme audio-visual processing, and inclusive, deterministic design for neurodivergent users, enabling sub-50ms responses on-device.
talk.summaryAiDisclaimer
Taha MountasserMount AI Scholar (Stealth Startup)
Traditional AI assistants rely heavily on cloud-hosted LLM APIs. While powerful,
this architecture poses two major barriers for assistive tech in education:
extreme latency (which disrupts active cognitive processing) and severe privacy
vulnerabilities for sensitive student data. For users with cognitive differences
like dyslexia or auditory processing disorders, a 2-second API response delay
is not just an inconvenience—it is a cognitive wall.
In this session, we break down the engineering of a privacy-first, zero-latency
assistive ecosystem running entirely on the edge. Using a real-world case study—
We will dive deep into:
1. Local Inference Optimization: Running Small Language Models (SLMs) like
Gemma on edge hardware with aggressive quantization, keeping latency sub-50ms.
2. The Phoneme-Grapheme Pipeline: How to build a highly parallelized
audio-visual processing layout with low-level DSP (Digital Signal Processing)
and custom speech-to-text feedback loops.
3. Architecting for Inclusivity: Designing deterministic, accessible application
states that accommodate neurodivergent working memory loads.
We will move past the hype of massive cloud-scale APIs to focus on what matters
for the next generation of software: offline-first, client-side intelligence
that treats privacy as a non-negotiable architectural boundary.
this architecture poses two major barriers for assistive tech in education:
extreme latency (which disrupts active cognitive processing) and severe privacy
vulnerabilities for sensitive student data. For users with cognitive differences
like dyslexia or auditory processing disorders, a 2-second API response delay
is not just an inconvenience—it is a cognitive wall.
In this session, we break down the engineering of a privacy-first, zero-latency
assistive ecosystem running entirely on the edge. Using a real-world case study—
We will dive deep into:
1. Local Inference Optimization: Running Small Language Models (SLMs) like
Gemma on edge hardware with aggressive quantization, keeping latency sub-50ms.
2. The Phoneme-Grapheme Pipeline: How to build a highly parallelized
audio-visual processing layout with low-level DSP (Digital Signal Processing)
and custom speech-to-text feedback loops.
3. Architecting for Inclusivity: Designing deterministic, accessible application
states that accommodate neurodivergent working memory loads.
We will move past the hype of massive cloud-scale APIs to focus on what matters
for the next generation of software: offline-first, client-side intelligence
that treats privacy as a non-negotiable architectural boundary.
Taha Mountasser
Taha is a 13-year-old developper and young architect from Morocco, building at the intersection of local Edge Inference and cognitive accessibility. As the creator of 'Mount AI Scholar', he designed a privacy-by-design local AI pipeline running low-latency SLMs (Gemma) at the edge to assist individuals with dyslexia. A competitor in Kaggle DeepMind hackathons and senior-level developer, Taha bypasses traditional academic timelines to focus on high-performance distributed systems, Swift/CoreML, and hardware-efficient machine learning. When he isn't benchmarking local inference models, he is preparing his stealth roadmap for the WWDC Student Challenge