John Jomi
Professional Journey

Experience

Internships where I researched on-device AI models and shipped production full-stack features, plus independent security research.

01Internship
Prophaze Technologies

AI Academic Intern

Researched small-footprint language models and hybrid reasoning architectures, benchmarking their viability for on-device and edge AI deployment.

  • Researched Small Language Models (SLMs) and Hybrid Reasoning Models (HRMs), evaluating their suitability for real-world AI applications.
  • Benchmarked on-device AI models across latency, memory usage, inference speed, and deployment feasibility for edge devices.
  • Performed comparative analysis of generative AI architectures through literature reviews and technical evaluations, documenting key findings for the research team.
  • Explored optimization strategies for mobile AI deployment, focusing on efficient inference in resource-constrained environments.
02Internship
Proddy AI

Full Stack Developer Intern

Shipped full-stack features for a modular AI work-management platform, spanning Next.js interfaces and REST API integration across multiple sprints.

  • Developed and enhanced full-stack web application features using modern Next.js, improving application functionality and user experience.
  • Built and integrated REST APIs, streamlining communication between frontend and backend modules for scalable development.
  • Collaborated with developers using Git-based workflows, participating in code reviews, bug fixes, and feature implementation across multiple sprints.
  • Optimized application performance by identifying and resolving UI and backend issues, improving overall responsiveness and maintainability.

Research

Independent Security Research

Self-directed work on phishing detection, pursued alongside coursework.

01ResearchOngoing
Tycoon 2FA — AiTM Phishing Detection

Independent Security Researcher

Designing a multi-layer detection pipeline for Tycoon 2FA and adversary-in-the-middle phishing email, built around an explainable signal model where every detection carries its own evidence.

  • Specified a four-layer detection architecture — header and domain intelligence, URL and redirect-chain analysis, NLP/ML body analysis, and threat-intelligence correlation — fused into a single composite risk verdict.
  • Designed a layer-independent DetectionSignal contract in which each signal emits human-readable evidence rather than a bare number, making per-layer ablation studies possible.
  • Scoped the prototype to read-only Gmail ingest of raw RFC-822 messages, taking no remediation action, to keep the research within ethical and legal boundaries.
  • Set a defensibility constraint for the project: every documented claim must trace back to a measurement produced by the evaluation harness.
  • Filed a copyright application covering the detection methodology; currently under review.