Gene Kastyshyn

I build AI systems that connect language models with real software, real interfaces, and real users.

AI & ML Engineer | Designing and leading GenAI solutions, agentic architectures, and real-world AI systems at scale.

I like AI most when it leaves the demo stage and starts doing useful work inside complex systems: reading screenshots, comparing metadata, explaining failures, interacting with web UIs, assisting developers, and helping non-experts operate software that would otherwise require deep technical knowledge.

My background is a mix of applied mathematics, full-stack engineering, product thinking, and many years of platform work. This combination shaped how I approach AI: not as a magic layer on top of software, but as a new computational interface that must be designed, measured, tested, and integrated carefully.

Current interests

I am especially interested in:

  • agentic AI systems that use tools, memory, planning, and evaluation
  • multimodal models that understand graphical user interfaces
  • AI-assisted coding beyond “vibes”: reliability, verification, and disciplined workflows
  • automated testing and troubleshooting with LLM/VLM agents
  • RAG systems grounded in private documents and metadata
  • practical AI product design inside large enterprise platforms
  • the boundary between software engineering, product management, and applied AI research

Much of my writing on this site is about these topics: how AI systems actually behave in day-to-day work, what current tools make possible, where they fail, and how engineers can collaborate with them more effectively.

Selected work

AI agents for test troubleshooting

In 2025 I built a proof-of-concept troubleshooting agent for ServiceNow Automated Test Framework failures. The system uses GPT-4o through Azure OpenAI and a LangGraph-based agent architecture to coordinate specialized agents for screenshot analysis, metadata comparison, diagnostics, and suggested fixes.

The goal is simple: when a test fails, the system should help explain why it failed and what a human should look at next.

Vision-language models for web UI understanding

I have explored how multimodal LLMs and Computer Use-style systems can interact with web pages in a human-like way: by looking at the screen instead of relying on DOM internals, element IDs, or implementation-specific hooks.

This is important for modern enterprise UIs, where the same user task may be implemented with many different front-end technologies. A useful AI assistant should understand the interface at the level of the user, not only at the level of the code.

Generative AI in automated testing

Shortly after ChatGPT became publicly available, I started experimenting with using LLMs for GUI understanding and test automation. Later work included GPT-4 Vision-based extensions to automated testing, allowing a test to validate or extract arbitrary visible information from a page.

Performance profiling and statistical comparison

I have worked on methods for comparing noisy performance measurements in automated tests and turning statistical results into product-level signals that non-specialists can understand. This work contributed to ServiceNow customer-facing functionality and patent activity.

Earlier platform and testing work

Before the current LLM wave, I worked on low-code/no-code web UI testing systems, including ServiceNow ATF Custom UI, GlideAutomateWait, and related patents and frameworks. A recurring theme in this work has been the same: reduce the amount of fragile manual knowledge required to test and operate complex software.

Background

I have a Master’s Degree in Applied Mathematics and 15+ years of experience in software engineering, SaaS platforms, product R&D, and AI/ML-oriented work.

Over the years I have worked as a Staff Software Engineer, Product Manager, Tech Lead, Automation Architect, founder, and full-stack developer. I have contributed to projects across ServiceNow, International Game Technology, Sears Holdings, eBay, and earlier startup work.

My technical toolkit includes Python, JavaScript, Java, ServiceNow Platform, LangGraph, LangChain, Azure OpenAI, OpenAI APIs, Anthropic Computer Use, Hugging Face Transformers, scikit-learn, pandas, NumPy, React, Angular, SQL/NoSQL databases, CI/CD tools, and the usual engineering glue that makes prototypes survive contact with reality.

I am also a Certified Scrum Master, have completed machine learning coursework, participated in Kaggle competitions, and have authored patents and RFCs around automated testing, performance measurement, and AI-assisted platform capabilities.

Writing

I write for engineers, AI practitioners, product-minded builders, and colleagues who are trying to understand what it means to work with AI as a practical partner.

Some recurring questions I care about:

  • What does it mean for an AI agent to “understand” a software task?
  • How do we make AI-assisted coding reliable instead of merely impressive?
  • What should be verified, measured, logged, and reviewed?
  • Where should humans stay in the loop?
  • How do LLMs change the interface between users, developers, and software platforms?
  • How do we turn new model capabilities into boring, useful, maintainable systems?

My goal is to write in a way that is technically grounded but readable: less hype, more mechanisms; less theater, more working systems.

Elsewhere

Outside of work, I like wingfoiling, kitesurfing, guitar, and occasionally staring at an AI agent wondering whether it forgot the task or discovered a better one.