Working with AI

How I use AI

One important thing about my use of AI is: I don't have to use it.

However, I do have experience using AI in multiple ways that can bring value and efficiency to your project:

• deployment, training, and content generation using AI image, video and audio models with node-based pipelines
• building AI agents that automate tasks and integrate with software
• generative coding using the latest techniques to speed up development

I cover all three below.

AI Content Generation

I work with generative AI across model training and content production:

• Model training

  Preparing datasets, pseudonymization, fine-tuning, and LoRA creation for character and environmental consistency, all on the fastest available GPUs. The assistant on this site is one example: a fine-tuned Qwen 2.5 trained on years of my own anonymized emails, served on RunPod and gated by Cloudflare.

• GPU provisioning and deployment

  Large-scale image and video work needs more than a chatbot. I provision cutting-edge GPUs, install models, manage checkpoints and LoRAs, and set up the right CUDA environment. Access to those GPUs is included in my hourly rate.

• Generative AI workflows

  Node-based pipelines (ComfyUI and similar tools) that get the most out of every GPU cycle: automating repetitive tasks, generating from a dataset, and creating subtle variations in each batch. These workflows are work-for-hire, so you own the copyright.

• Multimodal output

  Images, audio, and video with commercial rights, edited into a finished professional result. Custom music with human-written lyrics in any genre, mastered to a quality suitable for cinema or streaming platforms.

AI Agents

An AI agent does more than answer questions. It takes action. I build agents into apps and websites, and as standalone tools, to automate work that would otherwise eat up your time.

• Adding AI agents to apps

  Summarizing documents, classifying records, extracting structured data from messy text, and most things you might want automated. I keep API costs low with rate limiting and careful prompts. My own app, CalorieGoals, uses an AI agent to import recipes and auto-match foods to the USDA food database. It can also estimate portion sizes, exercise expenditure, and calorie counts from meal photos.

• AI chat agents

  A conversational agent that answers in your voice and takes action when asked, going well beyond a scripted FAQ bot. The chat agent on this site, MikeGPT, is a working example: it answers questions about my work, and if you ask, it takes your name and message and emails them to me. See how it works →

• Automating tasks

  Agents can run on a schedule or a trigger to handle repetitive work: watching for new data, processing incoming messages, generating reports, and moving information between systems. The busywork that quietly eats hours gets handed to an agent that does it the same way every time.

AI for Software

AI has changed how quickly software can be built, but only in the right hands. I use it where it genuinely saves time, and I skip it where it would cause problems. Either way, every line of code that reaches you is code I have read and understood myself.

When does AI help?

• Repetitive tasks

  Renaming things, converting data formats, updating an old library. AI does this kind of work in minutes, and it is easy for me to verify.

• Learning new systems quickly

  When a project involves a tool I have not used before, AI shortens the research phase considerably.

• Obscure error messages and logs

  AI is good at parsing dense information like debug logs, crash reports, terminal output, and raw API data.

• Well-defined functions

  When the input and output are known and constrained, AI is exceptional at writing the code to produce the correct output.

When do I avoid it?

I do not let AI make decisions about security, architecture, or what your project should do. Those calls require understanding your business, and they are the reason you hire an experienced developer in the first place. I also never ship code I cannot read and explain. AI output is a draft for me to review, not a finished product.

On smaller projects, or in well-understood open-source languages, I can give the AI more room to write the software itself. But once a project grows to medium or large, the AI tends to lose focus, burn far too many tokens just keeping track of the codebase, and slip into the occasional hallucination.

To use generative coding on larger projects without those problems, I rely on a few techniques:

• Phased planning

  The work is split into phases, each with a clear checkpoint.

• Automated testing

  Tests run after each phase, so nothing silently breaks along the way.

• Sub-agents and memory

  Code-review agents and persistent memory keep the AI focused on a large codebase.

• Central API

  One backend that the mobile, web, and desktop versions all reuse.

This is the difference between what people call "vibe coding" and using AI to ship a focused, production-ready app.