Building the Work Brain: A Glean Alternative in a Weekend
Self-hosted RAG over 45,349 Jira and Confluence documents for €60/month on AWS. Hybrid search, Reciprocal Rank Fusion, pgvector, and the three mistakes I made first.
·20 min read
AI Delivery
Production patterns for AI systems that actually ship. From single-agent orchestration to multi-agent coordination, RAG pipelines to MCP integration, these are lessons from real enterprise deployments.
AI delivery, as I use the term, refers to the engineering work between a working prototype and a system that runs in production with real users and real consequences. Most AI projects stall in that gap. The model performs well in a notebook, the demo impresses stakeholders, and then the team discovers that integration complexity, observability requirements, failure handling, and operational ownership were never planned for. The technology is rarely the bottleneck. The engineering around it is.
This series traces a deliberate progression. It starts with the organisational and process failures that sink AI initiatives before the architecture even matters. From there it moves into the mechanics of building a single production AI agent: orchestration patterns, tool design, guardrails, and the observability you cannot skip. RAG architecture follows, covering the retrieval pipeline that most tutorials gloss over, because generation quality depends almost entirely on retrieval quality. Multi-agent systems come next, along with a frank assessment of when you actually need more than one agent (less often than the conference talks suggest). The Model Context Protocol piece covers the emerging standard for connecting AI models to enterprise systems in a way that is portable, secure, and auditable. The series closes with a real-world case study: building a self-hosted knowledge search system over 45,000 documents as a practical alternative to enterprise SaaS tools.
Every article is grounded in systems I have built or helped teams ship across fintech, telecom, and gaming. Where I cite numbers, they come from specific deployments or published research. Where I offer opinions, I try to make the reasoning visible so you can disagree usefully.
AI systems depend on solid data foundations. See the Data Architecture series for the infrastructure layer.
Articles in this series
RAG Architecture That Actually Works in Production
Beyond the tutorials: practical RAG patterns for enterprise systems. Chunking strategies, retrieval architecture, and evaluation frameworks from real deployments.
·10 min read
Multi-Agent Systems: Patterns That Ship to Production
Architecture patterns, coordination strategies, and hard-won lessons from building multi-agent AI systems that actually work. Not theory; production patterns.
·9 min read
Model Context Protocol (MCP): A Practical Guide for Enterprise AI
What MCP is, why it matters for enterprise AI integration, and practical patterns for implementation. A no-hype guide to connecting AI models with your systems.
·23 min read
Building Production AI Agents: Lessons from Real Deployments
Moving beyond demos to production-ready AI agents. Architecture decisions, observability patterns, and hard-won lessons from deploying agents that actually work.
·26 min read
From POC to Production: Why Most AI Projects Fail (And How to Fix It)
The patterns that sink AI initiatives, and the engineering and organizational practices that lead to success. Lessons from projects that shipped and those that didn't.
·24 min read