Stop Prompting Your Agents to Behave: Use Sandboxing Instead
Prompts can shape what an AI agent tries to do. Sandboxing and capability controls determine what it can actually do.
Ndamulelo Nemakhavhani · AI · Cloud · Data
Systems worth understanding.
Field notes from building with artificial intelligence, cloud platforms, data systems, and the software connecting them.
Latest field note
Discussing the Trust Planes of Autonomous AI Work
How to establish trust in autonomous AI systems through identity, secrets management, and continuous assurance.
Read the essay →Recent writing
All 64 articlesWhy Desktop Apps Are Making a Comeback in the Age of AI
The browser won the last decade on reach. AI workloads are local, stateful and long-running, and the browser sandbox was never designed for that. The most serious AI tools of the last two years shipped as desktop apps, and this post argues that is not a coincidence.
Trading Expertise for Quick Answers - The Cost of Cheap Intelligence
Knowing something rare is no longer a moat. Most model answers are disposable by design. Expertise is what still stands after the next model ships: a source you can check, a habit of checking it, and residue that outlives the tab.
Evaluating Overlapping Plugins in Coding Agents
Overlapping plugins complicate attributing coding-agent outcomes to the model or its configuration. This post distinguishes package resolution from per-run behaviour and proposes controlled comparisons using invocation traces, correctness, safety and cost.
Different Types of Wikis and When to Use Them
Code repositories, documents, and agent instructions answer different questions about a software system. This guide maps generated repository wikis, codebase analysis, RAG, linked knowledge wikis, and the Open Knowledge Format to help teams choose the right documentation architecture.
Selected work
View all projects
Azure Batch Migration for Deduplication Workloads
I led the redesign and migration of a VM-based deduplication solution to Azure Batch. Moving the workload to managed batch compute improved scalability and reduced annual client infrastructure costs by approximately $8,000.
Language Models for Tshivenda and Bantu Languages
I owned the research and training of masked language models for Tshivenda and other indigenous Bantu languages. The work expanded language-technology coverage for languages underserved by mainstream NLP systems.
Regulatory Document Intelligence
I applied NLP and large language models to extract key information from regulatory documents. The workflow was designed to make review more structured, traceable, and less dependent on repetitive manual handling.