# bluenotebook.io

> Personal technical blog of Nikhil Kasukurthi, an ML engineer doing independent research in Göttingen, Germany, who previously led the AI team at Eka.care. Deep-dive articles on LLM infrastructure, AI agents, Model Context Protocol (MCP), cost-effective GPU training, and healthcare AI, with real cost breakdowns and production metrics.

Articles combine academic rigor with production engineering practicality: real dollar figures, benchmarks, profiling data, and honest trade-offs from systems serving 1M+ daily active users. Every post is available as plain markdown at the same URL with a .md extension.

## Posts

- [Does a context gate for search agents actually work?](https://bluenotebook.io/blog/context-gating-browsecomp.md): Testing context gating on BrowseComp-Plus. A cheap label-free gate cuts a search agent's input tokens 1.4x, gate costs included, no detectable accuracy change. (2026-07-17)
- [Which H100 instance to train Nanochat](https://bluenotebook.io/blog/h100-nanochat-training.md): Benchmarking H100 PCIe vs SXM vs NVL on training cost, step times, and NCCL profiling to find the cheapest GPU configuration for Nanochat (2026-03-04)
- [Why Model Context Protocol (MCP)?](https://bluenotebook.io/blog/why-mcp.md): Enabling endless capabilities for LLMs (2025-04-02)
- [Making LLM workflows human friendly](https://bluenotebook.io/blog/streaming-function-calling.md): Validate the data as it streams, don't make your users wait. (2025-01-28)

## Pages

- [About Nikhil Kasukurthi](https://bluenotebook.io/about/): Background, open-source work (KARMA LLM evaluation library, Eka MCP server), and publications
- [Resume](https://bluenotebook.io/nikhil-kasukurthi-resume.pdf): Full professional history (PDF)

## Optional

- [Full content](https://bluenotebook.io/llms-full.txt): All posts concatenated as one markdown file
- [RSS feed](https://bluenotebook.io/rss.xml): Subscribe for new posts