This repository contains projects that were discontinued or paused mid-way.
Although no longer under active development, these projects are kept for future reference or possible continuation.
The code inside these archived projects is often prototype-level and not fully polished.
They reflect experiments, partial implementations, or learning processes rather than production-ready code.
I chose to preserve them as-is, since they represent valuable learning experiences.
They may contain:
- Partial code implementations
- Experimental results
- Learnings from real-world constraints (e.g., lack of data, hardware limitations)
-
Description:
A time series and regression-based approach to forecast sales for small businesses, inspired by my family's chicken restaurant. -
Reason for Archival:
Although the project was nearly complete, I found that the dataset was too disconnected from real-world sales patterns.
Most available public data lacked granularity and had many missing values.
In reality, accurate modeling would require access to private credit card transaction records or point-of-sale data, which are not publicly available. -
Takeaways:
- Feature engineering was challenging due to missing values and data noise.
- Modeling can only be as good as the data — domain alignment is critical.
-
Description:
A semantic segmentation project on urban scene images, aimed at understanding road environments. -
Reason for Archival:
The project could not proceed due to hardware limitations — my local setup was not capable of handling high-resolution training at scale. -
Takeaways:
- Segmentation tasks are resource-intensive; scaling up requires either cloud computing or a powerful local GPU.
- Labeling complexity and memory bottlenecks are serious considerations in vision tasks.
-
Description:
An audio-feature-based recommendation system designed to suggest bass-heavy songs by analyzing low-frequency energy rather than relying on conventional metadata like genre or artist. -
Reason for Archival:
During development, I realized that the publicly accessible Spotify API imposes heavy limitations on access to detailed audio features.
Additionally, I lacked sufficient signal processing knowledge to implement my own robust bass-detection algorithm from raw waveform data.
As a result, the project was paused until either higher-resolution data access becomes available or my understanding of audio analysis improves. -
Takeaways:
- API access constraints can fundamentally limit ML system design.
- A deeper understanding of domain-specific algorithms (e.g., spectral analysis) is necessary before building a production-quality audio recommender.
While these projects are not active, they may still be:
- Picked up later with better data or hardware
- Referenced for code reuse or experimental structure
- Converted into lighter prototypes
📝 These projects represent learning moments — even when unfinished.