About

I'm Shiv Kumar. I work as a Senior Software Solutions Architect and a Senior Applied AI Engineer & Architect. I'm passionate about how systems are designed, how their intent is expressed, and how we establish that they do what we need them to do.

Software engineering

I care about clear responsibilities, meaningful abstractions, and code whose structure makes its purpose understandable. I care just as much about verification: defining the expected behavior and writing functional acceptance tests that exercise it at the system boundary.

Programming With Intent brings much of that thinking together. I've put years into making design and testing guidelines thorough and specific enough to apply consistently. That work has new relevance as AI takes on more of the implementation and maintenance of code. I want those systems to work against explicit engineering standards, with tests that demonstrate whether the software meets its requirements. Human judgment defines the intent and the standard of correctness.

Applied AI engineering

Much of my applied AI work involves document intelligence: extracting structure and metadata, identifying requirements, and turning information spread across documents into strategic intelligence people can use. I build pipelines that combine model interpretation with processing and validation in code. I also explore conversational systems that let people question their documents and investigate relationships across them.

Controlled workflows

One problem I keep returning to is how to get greater determinism and repeatability from systems built around language models. My approach is to build directed graphs whose controllers and orchestrators are imperative code. Code manages dependencies, branching, validation, and failure handling. I use models where interpretation, judgment, or generation is required, and keep exact operations in code. This is how I approach complex document pipelines and software verification workflows.

My work with mixtures of experts assigns extraction, independent review, diagnosis, and correction to separate specialists. Each has a defined responsibility and a constrained view of the evidence. Model judgments can still vary; explicit control over execution and acceptance makes the system more predictable and its results easier to inspect.

Decision models such as Jevextend that approach. They make fast, bounded judgments about text and return structured decisions and probabilities that code can act on. That gives a controller access to semantic judgments that would be difficult to express as fixed rules, while keeping control of the workflow in code.

Agentic systems

I'm equally interested in systems where agents have room to decide how to pursue a goal. Anthea, my personal chief-of-staff project, is an important part of that work. Its thesis-defense system adapts to the position someone submits. It examines the thesis for gaps, assumptions, and leaps between categories of claims, researches the wider subject, and creates specialist agents suited to challenge it.

Each agent has its own persona, perspective, directive, and goal. Within its role, it can use tools, seek evidence on the internet, and develop a sustained attack on the argument. The opposition takes shape around the thesis being examined. Building and working with that system has taught me a great deal about my own thinking: the assumptions I leave unstated, how I respond to objections, and what it takes to defend a position under serious scrutiny.

What I share here

This site brings those interests together through articles, worked examples, and material from my videos. I explain the reasoning behind design choices, the problems they address, and where they apply. Thewriting covers software engineering and applied AI;Programming With Intent andAcceptance Testing collect the detailed engineering guidance, with examples in Python and C#.

How I produce articles

Every article starts with my own experience: experiments I run, problems I encounter at work, or ideas I explore in my YouTube videos. I write code to explain and demonstrate concepts, test ideas, and establish what supports a conclusion. Sometimes that work becomes both a video and an article; sometimes it becomes an article on its own.

When I can share the example code, it is available on my GitHub. When the experience comes from work I cannot share, I build an independent example that demonstrates the idea without exposing company, product, or proprietary implementation details. The article gives readers the explanation and evidence they need to follow the reasoning.

I draft the article from that experience and my own reasoning. AI comes in afterward, helping me clean up the draft and refine its clarity, structure, and wording. The writing follows the style I have developed over the years. Phrases such as "verified, not trusted" and "instruction is not assurance" come from my own thinking, as do the ideas, examples, and engineering judgments behind the article.

My article workflow includes a separate review of teaching flow, checks of claims and citations against source evidence, and revision followed by an independent comparison with the earlier draft. I review the result and am responsible for what I publish.