Hi there, not a very unique post I am sure, since many of us are beginners trying to build portfolio and extend our knowledge both width and depthwise.
I have personally started (or got involved in) a number of mini-projects from different ML domains (all of which are ongoing WIP cause life is full of distractions), and I would be happy to collaborate on some of them, hear ideas for alternative approaches, learn from each other etc.
I’d aim for people who are beginners like me, but with some solid foundations and intuition for basic ML concepts and who also have some projects behind them, so we can motivate and learn from each other, but I am also happy to help out with more basic problems if I am able to.
This is a very long read so here is a TL;DR chatGPT summary for these with insufficient time or attention span 
- Reinforcement Learning repo (Gymnasium-based): Implementing core RL algorithms (e.g., Deep Q, A2C, REINFORCE) from scratch on discrete and continuous environments, including training Deep Q to play Breakout, with documented debugging insights and plans to expand toward more advanced environments.
- Unsupervised Learning on EEG data: Pretraining models on large-scale unlabeled 128-channel EEG time-series using contrastive learning, followed by supervised fine-tuning to predict response times; plans to extend with VAE, masked reconstruction, and latent space visualization (UMAP, t-SNE).
- Audio Classification & MIDI Transcription: Detecting drum hits (kick, snare, hi-hats) and transcribing audio into MIDI using STFT and temporal CNNs; future extension to pitch detection, instrument classification, and full melodic transcription—motivated by music production in Ableton Live.
- RAG-based Job Matching Pipeline: Building a lightweight RAG system using LLMs and Chroma to scrape job listings, embed and store them in a vector DB, and rank opportunities based on CV similarity; currently functional for one ML job site with plans to expand via APIs.
- Anomaly Detection for Predictive Maintenance: Exploring anomaly detection on aerospace flight-phase data and experimental IoT data (e.g., IMU sensor on espresso machine) for malfunction detection and predictive maintenance use cases.
- Computer Vision for Posture Evaluation: Developing a CV-based system to analyze exercise/kickboxing posture from video and detect common mistakes in movements like kicks and jabs.
If you wish to know more …
Existing GitHub repos:
Below is a list of topics I cover in my GitHub, as well as some ideas for future projects. Perhaps you will find some of it interesting:
-
Reinforcement learning repo: supposed to contain different reinforcement learning algorithms, mainly trained on Gymnasium type environments.
Currently quite limited, but I plan to expand it and would be happy to learn RL together.
Idea is to write basic RL algorithm such as DeepQ, A2C, REINFORCE etc on number of discrete and continuous environments and play around with parameters to gain better insights into how things work for these who are just starting, before moving on to more ready made solution, such as Isaac Gym/Lab.One cool entry project was getting DeepQ algorithm to play classic Atari game Breakout. I ran into some issues along the way and I documented it in Medium article documenting the debugging process (with the help of the users here). Hope it may be insightful for others.
-
Unsupervised learning on EEG data. I recently entered a competition in which the task was utilizing vast amount of unlabeled EEG data to pretrain the model using unsupervised learning to extract meaningful features from the 128 channel timeseries data (I personally went for contrastive learning approach).
Model is then trained using supervised learning on much scarcer labeled dataset to predict response time of subjects in performing some detection and reaction tasks.
Goal of the first part was to show that supervised learning has helped model abstract some basic features out of the signals and learn that similar signals belong closer together in the representation space.
Results indicate that this was achieved, however I want to further improve pre-training with VAE (variational autoencoder), masked reconstruction etc. Also I would like to make some cool visualization using UMAP and t-SNE to show how latent space changes during pre-training
Its a great project to learn more about unsupervised learning. -
Audio classification and transcription into MIDI.
I am currently writting a ReadMe file for this one (hopefully ready in a day or two), as its the project I am mainly focused on atm.
Idea is to analyse STFT - short sliding windows of frequency content in an audio stream and trying to detect individual drum hits (e.g. Kick, Snare, Hi-hats etc) and automatically transform it into MIDI.
Second part would be to recognize individual notes (and perhaps even classify instruments) and transcribe into midi as notes of certain pitch e.g. Piano track: C,D,E…
Its an interesting problem of multilabel-classification on timeseries and it utilizes a lot of interesting concept, such as temporal CNNs, data augmentation etc. There are also many different approaches one can try here, and the drum detection is currently slowly shaping up to actually give some sensible results
Motivation for this project is a bit silly - I sometimes write music in Ableton and when I have and idea on the go, I would like to just hum, whistle or beatbox it into my phone and then have a draft of an idea as a midi file I can import directly in the software.
-
I worked a bit on RAG and generally utilizing LLM’s to work on processing data.
More specifically, I made a basic * RAG-like pipeline to fetch and analyse job data from various sources (e.g. job sites), process them to extract basic information using lightweight language models, store the data in a vector DB using Chroma, which can be easily queried based on similarity, and then further processed based on my given CV.
This way it categorizes jobs from multiple sources based on how well they fit my CV and desires. As everything else - still very much WIP, tho main pipeline is in place and processes data from one ML jobs website, tho I wish to migrate to APIs that cover larger job markets.
Some further ideas I didn’t yet manage to get into:
-
Anomaly detection for predictive maintenance /action - I am working in a job where we collect data from flying things, so I have a lot of data from various flight phases available, where I would like to test some anomaly detection
I also hooked up a basic IMU sensor to a microcontroller to record some data from my espresso machine to check for malfunctions or detect time for refill, but never got around to it. (was mostly excited to work with micro-controllers again)
-
Posture evaluation - I used some of the existing CV libraries to analyze posture during exercise. As I am doing kickboxing I was thinking about making an app to point out common mistakes I do when e.g. practicing kicks and jabs, since they are quite easy to detect in video.
A little bit more about me
If you made it this far, here is a few more extra notes about me:
I come from a control system background and have worked mostly in aerospace and aviation, but decided to go more in ML direction lately. This makes me bit more biased towards system identification, time-series analysis tho my interests go quite beyond it. Control System background gave me some solid background in linear algebra, statistics, calculus and other math relevant for understanding ML concepts.
Thanks for reading thru this behemoth of text and hope to hear your project ideas.