Software DIY in the age of AI

DIY In The Age Of AI

The other day I was looking for a translation tool that could handle a large number of documents. As I searched through various tools, I found that the good ones were expensive, and the cheaper ones didn't have the features and language packs I needed. Naturally, I started exploring Python libraries and LLMs to see if I could build my own Local LLM. With help from AI assistants in my IDE, I ended up developing a very custom tool tailor-made for my needs. It didn't have many of the bells and whistles that commercially available tools had! And in reality, I rarely use all the features that products provide. I used free models in Ollama, and the translation quality was quite good! Moreover, I was able to switch models based on the language - Qwen for Chinese and Arabic, Llama3 for all other languages. I played around with parameters like generation temperature, chunking, and tokens to dramatically improve the translation quality. And in no time, I had a good-quality translation engine that stayed on my machine and did not consume AI tokens! When you buy commercially available tools, then all this plumbing is already done for you. But if you want to get your hands dirty, you can play around with this and get free software. Of course, you would be paying for AI code-gen tools. Should I be looking at building all the software that I use myself? It will be purpose-built, free, and flexible. There is always an argument that paid software will have a roadmap and continuous improvement. But if I can use AI tools to create a roadmap when I need it, tailored to what I need, it actually makes more sense to build it myself.

I used the same logic to build a project management software. I did not want all the rich features to analyze a project in 30 different ways. All I needed was to define a simple project with dates, resources, and dependencies, and a bunch of agents to follow up and adjust the timelines based on inputs. Works beautifully and is free. Increasingly, I am becoming a fan of using Ollama for production deployments and paid tools like Claude Code and PyCharm to build the tools. We still need to buy some software. The rest can be built.

For large enterprises, building all tools to save cost would not make sense. But for smaller startups that are cost-sensitive and individual users, building your own might be a good way forward.