For my ml project, I started with two ideas:

- A baldness/makeup classifier 
- An Ethiopian music genre classifier 

Both ran into the same problem: data.

Finding a quality dataset for the makeup project was nearly impossible, and since i don't know enough about music manually labeling Ethiopian music genres would have taken forever.

after of getting stuck I pivoted.

I used a movie API to collect data, downloaded hundreds of movie posters, and built an movie image genre classifier that predicts movie genres from poster images alone.

The training process was full of trial and error:

- 📉 Epoch 1: High validation loss — the model was underfitting. 
  🔧 Fix: Standardized image dimensions and preprocessing across the dataset so the model could learn meaningful patterns.

- 📈 Epoch 10: Validation performance started dropping — classic overfitting. The model was memorizing posters instead of generalizing.

- ✅ Epoch 3: The sweet spot. Balanced training and validation loss with solid performance metrics.

One thing this project taught me:

> Machine Learning isn't just about writing code. It's about collecting data, cleaning it, experimenting with models, tuning hyperparameters, analyzing results, and repeating the process until everything clicks.

The model isn't always the hardest part. Sometimes the real challenge is the data itself.

Lesson 2 was all about deploying a model to production exporting the final .pkl file after training and finally deploying it on Hugging Face unfortunately I had a hard time doing that because I encountered errors from gradio on the huggingface when passing image data to the model so I left that part for now and just ran it in Google Colab instead I tested it there and it was quite interesting.

Do it or do not do it either way you will regret both.

Soren Kierkegaard

hey if you are like me and use a lot of CLI AI tools right now you probably know this pain. lately whenever i want the agent to fix something or i need to explain a complex architecture in detail and discuss it i find myself completely lost in the middle of typing. like i literally lose my train of thought and forget what i was even gonna ask or prompt the agent by the time i'm halfway through typing it out lol
​so i was thinking it would be great if we could just speak and voila the prompt gets sent to the AI. most apps do that now but the CLI doesn't really have a built in way to do it. or at least i didn't know how until now.
​turns out you don't need to build anything. you just place your cursor right in the terminal and press the windows key + H at the exact same time. it pulls up the built in windows voice tool and you can just talk and it types everything right there. saved me so much headache today thought i should share.

hey if you are like me and use a lot of CLI AI tools right now you probably know this pain. lately whenever i want the agent to fix something or i need to explain a complex architecture in detail and discuss it i find myself completely lost in the middle of typing. like i literally lose my train of thought and forget what i was even gonna ask or prompt the agent by the time i'm halfway through typing it out lol
​so i was thinking it would be great if we could just speak and voila the prompt gets sent to the AI. most apps do that now but the CLI doesn't really have a built in way to do it. or at least i didn't know how until now.
​turns out you don't need to build anything. you just place your cursor right in the terminal and press the windows key + H at the exact same time. it pulls up the built in windows voice tool and you can just talk and it types everything right there. saved me so much headache today thought i should share.

Been trying the free Buffer cli with DeepSeek V4 lately, and honestly it's been pretty good so far.

One thing I really like is the amount of tokens they give you. It's actually generous. U can do a ton of projects task with it without constantly thinking about running out of credits or limits, which is something I appreciate a lot.

So far, from the testing I've done, I've found it genuinely useful.

The only thing I'm kind of disappointed about is the permission system. You know how coding cli ai agents tools let you switch modes? Like, one mode where the agent asks for permission before writing or running scripts , and another where it can just go ahead and execute everything automatically.

In this case, you have to explicitly tell it upfront that you want a plan first. And to be fair, it does give you the plan. But after the third or fourth iteration, it kind of starts assuming that you agree with what it's doing and just continues from there also they only allow one session.

That's probably the only catchs I've found so far.

Other than that, honestly, it's been really good. The token limit alone makes it worth trying if you're building projects or experimenting with ideas.

Has anyone else been using it? Curious to hear what your experience has been.

The way he sleeps makes me think he forgot how to be a real cat #skinwalker