Diego J.'s Software Tar Pit

I have created a C compiler that has improved features on safety and defined behavior. I am not an expert in Rust, not in compiler construction. How have I done it?

Introduction

In this post I am going to show how it is possible to create software where your knowledge of the domain and my Rust programming language skill are both limited. I will show how with a good spec, a validation process, and a Large Language Model (LLM) it can be done.

LLMs (Large Language Models) provide a wonderful tool for the construction of software that speed up the process of coding. But are there some downsides to them?

Introduction

The past 3 years have been a total change in the software engineering discipline, first the AI-based autocomplete, then the LLM-based code generation. Now, non-software engineers are creating and releasing software in the wild, and software engineers are being pressured to produce software faster.

We often hear that Large Language Models (LLMs) empower engineers by augmenting their existing expertise. But what happens when the engineer is a complete novice in the domain? Can a LLM be more than just a powerful autocomplete? Can it be a partner when developing for a domain you have never touched?

Introduction

I have been “vibe coding” extensively for the last six months. My background is primarily in web technologies, but I decided to push the limits: build a desktop application in a language I have played around but don’t consider my “forte”. So I decided to use golang to develop a videogame.

The cloud did not removed complexity, it only centralized it.

The infrastructure became more programmable, flexible, scalable, and abstracted. But that did not mean that the systems became simpler.

Architecting for the cloud

The cloud companies were born as a way to reduce complexity for the engineering teams. However, the complexity was not removed, but moved from the engineering teams to the cloud systems.

But creating complex architectures in the cloud has a clear counterpart: how can we ensure they work properly? How can we make sure that our solution does the right thing, and is sufficiently performant?

I have been using Gemini for a while and I was up for a challenge of creating a complex project with microservices for testing their capabilities, so I created a scalable scraper.

Introduction

Some years ago I worked at a company that had a scraper. The architecture was already in place and I had to fix scalability issues, release it to AWS, and create a user interface.

The other day I thought about using LLMs to replicate the project and talk about the experience. So I have been working with Gemini, and I have I created a scraper (Isidorus Web Scraper) that could run in AWS with minimal configuration and that would perform well.

Have you wondered what to do when a feature takes more time than the scheduled one? What can you do if you reach the deadline but the feature is incomplete?

Introduction

In this post I will tell an early in my career story when I first learned that the release of anything is always better than release nothing.

The contract

There was a contact where the company I worked for at that time, needed to provide a web service to a local branch of one of the most important companies in the world of c…¹.

In this post we pick up a Python 2.0 project and show how to use a LLM to upgrade a software project. This time we will be using contexts.

Introduction

I will upgrade the Djanban project as it should be done. Not only making it work with the latest versions of the software packages, but improving its quality and making sure there are guardrails that make impossible for the developers

The problem

As explained in the first post of this series, the backend of Djanban was implemented in Python 2.0 and Django 1.11. Python 2 has been outdated for 6 years now and is not used anymore. Django is now in version 6.0.1.

In this post we are going to lay some guidelines about how to prevent your project of accumulating tech debt.

Introduction

Tech debt has been a constant in all my work. All projects I have started working with had an amount of tech debt. Why does this happen? And better yet, what can we do to avoid falling on this tech debt trap?

What is tech debt?

Tech debt was a metaphor created by Ward Cunningham to help non-technical people the necessity of investing resources in keeping the software projects with the lesser level of accidental complexity possible.

In this post we pick up a Python 2.0 project and show how to use a LLM to upgrade a software project.

Introduction

I will upgrade the Djanban project as it should be done. Not only making it work with the latest versions of the software packages, but improving its quality and making sure there are guardrails that make impossible for the developers

The problem

As explained in the previous post, the backend of Djanban was implemented in Python 2.0 and Django 1.11. Python 2 has been outdated for 6 years now and is not used anymore. Django is now in version 6.0.1.

In this post we pick up a Python 2 project and show how NOT to use a LLM to upgrade an outdated software project.

This is a two-series go to the next post if you are only interested in how to do it correctly.

Introduction

We tend to only focus on positive outcomes, following active (positive) advice, etc. But what about negative one? I wanted the showcase an experience about upgrading a project using LLMs but not done it correctly. With so much advice on how to do things, why couldn’t we share how not to do it?