# Jonathan Barazany

> Chief AI at Nayax sharing insights on AI agents, data engineering, and technical leadership in enterprise software development.

## About the Author

Jonathan Barazany is Chief AI at Nayax, where he leads AI agents and R&D. He previously spent 10 years at Microsoft, where he led the adoption of AI coding agents. He specializes in AI agents, data engineering, distributed systems, and technical leadership, and creates practices used by thousands of developers.

- Website: https://barazany.dev
- LinkedIn: https://www.linkedin.com/in/jonathan-barazany-14925849/
- GitHub: https://github.com/johnib
- Twitter: https://twitter.com/barazany

## Site Sections

- [Blog](https://barazany.dev/blog): Articles on AI agents, data engineering, and technical leadership
- [About](https://barazany.dev/about): Background, experience, and current work
- [Projects](https://barazany.dev/projects): Open-source tools and side projects
- [Contact](https://barazany.dev/contact): Ways to get in touch

## Blog Posts

- [I Almost Bought an AI Code Reviewer. Then I Did the Math.](https://barazany.dev/blog/claude-code-pr-reviewer): Eight months of running automated AI code review on my own projects. Then I brought it to Nayax — and didn't buy a single dedicated tool to do it.
- [I Had Claude Mine 30 Days of My Own Messages. What It Found Was Uncomfortably Accurate.](https://barazany.dev/blog/mining-claude-history-ultracode): For months I kept correcting Claude the same ways. Instead of correcting it again, I asked it to go read everything I'd ever said — and tell me who I am.
- [The 5-Hour Quota, Boris's Tweet, and What the Source Code Actually Reveals](https://barazany.dev/blog/claude-code-token-caching): Everyone's complaining about the Claude Code quota. Boris said improvements are coming. My first reaction: what can he possibly do? Then the source code dropped.
- [Claude Code's Compaction Engine: What the Source Code Actually Reveals](https://barazany.dev/blog/claude-codes-compaction-engine): A few months ago I described context engineering from the outside. Then Claude Code's repository surfaced - and I asked Claude to analyze its own source code. Here's what's actually running beneath every long session.
- [We Were About to Buy an Automation Platform. We Already Had One.](https://barazany.dev/blog/we-already-had-the-automation-platform): We spent weeks in governance spirals over n8n and Workato. The answer was already deployed — it just needed one more feature to click into place.
- [What Karpathy's Autoresearch Unlocked for Me](https://barazany.dev/blog/what-karpathys-autoresearch-unlocked-for-me): I was stuck at 0.58 AUC for three weeks. Then an AI agent ran 165 experiments while I slept — and broke through a ceiling I couldn't crack on my own.
- [How We Measured 65% AI-Written Code (And Why Lines Don't Matter)](https://barazany.dev/blog/how-we-measured-65-percent-ai-written-code-and-why-lines-dont-matter): After our Claude Code pilot hit 65% AI-written code, the most common question was: how did you actually measure that? The answer is simpler than most expect - and it starts with abandoning the wrong metric.
- [What 65% AI-Written Code Taught Us About 'Good Enough'](https://barazany.dev/blog/what-65-percent-ai-written-code-taught-us-about-good-enough): A pilot group of 30-40 engineers set a 40% AI-written code target. Six weeks later, the real number was closer to 80%. What we learned about adoption, power users, and the shift to 'good enough.'
- [Connecting Snowflake to Claude: The Guide That Should Have Existed](https://barazany.dev/blog/connecting-snowflake-to-claude): It took a month to connect Snowflake to Claude. Not because the setup is impossible — but because the right documentation doesn't exist. Here's the guide I wish I had.
- [Good Enough and Really Fast: A Haiku 4.5 Case Study](https://barazany.dev/blog/good-enough-and-really-fast-haiku-case-study): A development team moves so fast that Product can't keep up. How a shift to Haiku 4.5 changed their economics and workflow, and what it reveals about the future of coding agents.
- [Cursor 2.0: When Your Code Editor Becomes a Work OS](https://barazany.dev/blog/cursor-2-0-work-os): Cursor 2.0 reimagines code editors as orchestration platforms. Exploring multi-agent workflows, parallelization trade-offs, and how AI tools are fundamentally changing how we think about work.
- [Haiku 4.5: The Model Nobody Expected to Care About](https://barazany.dev/blog/haiku-45-the-model-nobody-expected-to-care-about): A small, fast, affordable model that matches Sonnet 4-level capabilities. The gap between what we think we need and what actually gets the job done is closing faster than most people realize.
- [Z.ai and the GLM-4.6 Model: Good Progress, But Not Sonnet-Level Yet](https://barazany.dev/blog/glm-46-china-ai-breakthrough): GLM-4.6 is gaining attention as a capable, affordable alternative to Claude, but after nine months with Sonnet models, it's solid progress without redefining the frontier.
- [Context Engineering: What Keeps AI Agents From Losing Their Minds](https://barazany.dev/blog/context-engineering-what-keeps-ai-agents-from-losing-their-minds): Behind every long-running AI agent, there's quiet, invisible engineering that keeps the model from forgetting who it is and what it's doing.
- [Anthropic vs OpenAI: The AI Model Rally That Changed Everything](https://barazany.dev/blog/anthropic-vs-openai-ai-model-rally): A week of rapid-fire AI releases changed the game. GPT-5 challenged Claude's dominance, Anthropic countered with 1M context windows, and Claude 4.5 followed. Here's what this means for developers.
- [How MCP Prompts Automated 600-Line Kusto Query Debugging](https://barazany.dev/blog/how-mcp-prompts-automated-kusto-query-debugging): From manual pain to automated precision: how I used MCP tools to turn the nightmare of debugging complex Kusto queries into structured, data-driven performance reports.
- [AI Agents Debug Spark Faster](https://barazany.dev/blog/ai-agents-debug-spark-faster): How I turned a painful Spark debugging cycle into a tight feedback loop using Databricks CLI, DuckDB, and local iteration. From cluster slog to terminal efficiency.
- [Why I Set Kubernetes CPU Requests Equal to Limits (And You Should Too)](https://barazany.dev/blog/optimizing-resource-allocation-in-kubernetes): A production mystery that changed everything I thought I knew about Kubernetes resource management. When our 8-core workload couldn't get 8 cores, I discovered the real problem with traditional request/limit configurations.

## Topics Covered

- AI coding agents and development workflows
- Apache Spark debugging and optimization
- Model Context Protocol (MCP) integrations
- Databricks and cloud data pipelines
- Engineering leadership and team practices
- AI model comparisons and evaluations
- Kubernetes resource optimization
- Context engineering for AI systems