Applied AI Toolkit
Hi, I'm Federico Medinilla, a software developer in Buenos Aires, Argentina. I spent six years building websites and front ends. Since 2024 I have run KyndredAI, a small studio where I build AI assistants, n8n automations and web apps for schools, shops and small businesses.
This repository is the toolbox behind that work. When a client project needed something I knew I would need again, I pulled it out, wrote tests for it and put it here. You will find the code that keeps Claude API spending under a cap and the runner I use to measure a prompt change before it ships. Next to them are the security checks I run on the sites I build, and eight small apps that show the pieces working together. It is open source, so you can download any part and use it in your own projects.
How this repository is organized
The parts put an LLM into an application under a spending cap and measure a prompt change on the real model before it ships. They check what the model writes before a visitor sees it, and they harden the sites and databases around it. Each library folder has its own README and tests, and uses another folder only through its public entry (content-forge uses gateway). Each app under apps/ uses libraries only through their public entry. Every folder's README gives the design decisions with the reason for each, and its tests run offline with no API key. CI runs all of them on every push to main and every pull request.
Where to start
Live demos
The apps that work as static files are published on GitHub Pages at https://fedemedinilla.github.io/applied-ai-toolkit/, built and deployed by .github/workflows/pages.yml on every push to main. Nothing typed in them leaves the device.
| Demo | Try this |
|---|
| Caja | Type a sale in Spanish and answer its question when the amount has two readings. On Pages the optional model mode stays hidden, because it needs the app's server. |
| Receipt | Load the order with emoji and right-to-left names and download its PDF. |
| Replay Lab | Export a run, change one value in the file, and import it to see both hashes fail. |
| Site checklist | Mark a check Fails, then try to mark it Fixed before acknowledging the pre-change list. |
The chat widget, the writer, the salon assistant and push-lab need their own server, so they are not published; each README says how to run it. GitHub Pages cannot send response headers, so the published apps run without the Content-Security-Policy that their csp-guard.config.json checks the build against. To publish from a fork, enable Pages once in the repository: Settings > Pages > Source: GitHub Actions.
Apps
Each app builds with npm run build:<app> into dist/apps/<app>/. Caja, Receipt, the writer, the chat widget, push-lab and the salon assistant have Spanish interfaces; Replay Lab and the site checklist are in English.
This is the beginning of the README. Read the rest at the original. Read the full README ↗
