About:

Luke Marsden is the author of HelixML, a Substack publication that brings the best of open source AI to businesses. The publication has thousands of subscribers.

Website:

Specializations:

Incoming Links:

Subscribe to RSS:
The post outlines the challenges and solutions in forking the Zed code editor to enable a fleet of headless coding agents for Helix, focusing on performance and bug fixes.
Helix replaced WebRTC with WebSockets for real-time video streaming, resulting in lower latency and simpler infrastructure for their AI coding agents.
Evaluating AI coding assistants is complex due to inadequate benchmarks that fail to reflect real-world coding practices and the need for more comprehensive assessment methods.
Helix Desktop allows multiple isolated Linux environments on Apple Silicon Macs, enabling efficient GPU sharing and real-time collaboration between AI agents and human users.
The post discusses the development of interactive desktop environments for AI agents at Helix, utilizing the Moonlight game streaming protocol. It highlights the challenges of adapting a single-user gaming protocol for multi-user ...
The author shares the journey of replacing a complex video streaming solution with a simpler, more reliable JPEG screenshot method for their AI platform, Helix.
Helix revolutionizes software development by deploying multiple AI agents in isolated environments, enabling efficient parallel task execution and improved project management.
The migration of Kodit from Python to Go improved performance and usability, revealing critical lessons about architecture and the limitations of AI in coding assistance.
Helix's innovative caching system reduces Docker build times from 45 minutes to just 14 seconds, enhancing efficiency for development agents.
The author discusses advancements in AI coding agents, highlighting the limitations of current models like GPT-5 and Claude 4.5. They propose a new architecture where agents run on servers rather than local machines, allowing for ...
The article discusses the dynamic behavior of agents using Large Language Models (LLMs) and how to implement a 'relaxed' plumbing system for integrating multiple remote MCP (Managed Cloud Platform) servers. It covers the extractio...
The blog post discusses the release of Kodit 0.5, highlighting significant changes made to the software's architecture, including a shift to a Git-based domain model for better incremental indexing and database restructuring. The ...
Helix's integration with SharePoint and Teams enables AI agents to provide instant answers using existing company documents, streamlining enterprise communication and knowledge access.
Kodit has reached a significant milestone with version 0.3, featuring a 10x improvement in indexing speed through smarter batching and streaming generators. Key enhancements include private Azure DevOps support, pre-filter search ...
The blog post discusses the launch of Kodit 0.4, a significant update to the AI coding assistant that now includes a hosted service for accessing public repositories without installation. Key features include incremental indexing,...
The post discusses the implementation of MCP (Multi-Channel Protocol) authentication, highlighting its initial simplicity and the misconceptions surrounding its complexity. It details the process of integrating OAuth support in He...
Priya Samuel announces her new role as Head of Engineering at HelixML, emphasizing the importance of software craftsmanship and a positive company culture. She reflects on her experiences with AI implementation challenges and her ...
The post discusses the launch of Helix 2.0, a private GenAI stack developed by a startup that has recently achieved $1M in annual revenue. The author reflects on the journey of building the product, the skepticism around AI, and t...
The text discusses the limitations of AI coding assistants due to data blind spots and introduces Kodit, an MCP server that indexes codebases and offers relevant code snippets to coding assistants to overcome these limitations.
The text discusses the blind spot in AI technology that makes it unable to understand diagrams, images, charts, and tables in documents. It introduces VisionRAG, a new technology that can extract exact numbers from charts and unde...
The text discusses the issue of GPU hoarding, where many teams have high-performance GPUs sitting unused in data centers. The author promotes using HelixML to put these dormant GPUs to work, providing immediate access to a secure,...
The text discusses the unexpected benefits of self-hosting Large Language Models (LLMs) on your own Kubernetes cluster. It emphasizes the advantages of self-hosting, such as security, data protection, and data sovereignty, and the...
The author discusses the importance of prompt engineering in AI applications and the need for proper versioning and source control. They emphasize the need for treating prompts as software engineering challenges and the importance...
The text discusses the author's journey in building a call analysis assistant using voice data and the Helix AI application platform. The author shares the challenges faced, the discovery of the VCON format, and the benefits of us...