Add AI features to your product

Where AI earns its place

The question I ask first is simple: what job does a person do by hand today, and could a model do part of it well enough to help? If the answer is no, I say so. AI is a good fit for questions over a body of text, for turning rough notes into a structured result, and for summaries. It is a poor fit where an exact answer is required every time.

For an existing system, AI is added on top of what you have. For a new product, it is designed in from the start. My Services page has the other four services.

Examples from my own products

These are real products of mine, and each one has limits.

• Sidekick Note. Private Markdown notes that turn into cards: recipes, checklists, charts, converters. An AI tab can produce a card, such as a vegan recipe card or a unit-convert card, from a plain request. The demo shows real runs of the built-in AI.
• อปท.com. A directory of 7,800+ Thai local-government offices, with an AI question assistant.
• Warroom. My own daily-use terminal app has an AI read of the market state. It is not investment advice, and no performance is claimed.

All screens in the demo are sample data or my own notes.

One gateway behind several products

I run one shared AI gateway behind several of my products, with per-person credits. A new product can add AI without rebuilding the plumbing: keys, limits and usage tracking are in one place. It uses Ollama with a Node and TypeScript service. Read the IDUnify part of the platforms page for the identity side.

Prompts, context and limits

Most of the quality comes from the context you give the model, not the model name. I design the prompt and the context together: what the model sees, what it must refuse, and how an answer is shown to the user with a warning where it can be wrong.

Cost, limits and safety

I describe these in general terms, because they depend on your product.

• Cost follows the model you choose, the size of each request and how many people use it. I name these drivers in the consult.
• Limits per person and per day protect your bill. A credit system is one way to do it.
• Safety means the AI sees only the data it needs, answers are marked as AI, and users can see when it may be wrong.

What I do not promise

I do not promise that the AI is always right. I do not promise a cost saving or a number of users. I do not promise a specific model will stay available. I do promise a written scope, tests, and a plain note on the limits in the product itself.

See also

Read about how I work or see the products on the demo page.

Questions