ThinkFlow

Problem

EEG signals are noisy, sparse, and hard to align with language. The project explored whether learned discrete representations can support a text-generation pipeline.

System Shape

1. 01Preprocess raw EEG signals into model-ready temporal features.
2. 02Train representation models to compress noisy signal patterns.
3. 03Use sequence models to map learned representations toward language tokens.
4. 04Serve inference through a lightweight Flask and Next.js application.

Highlights

• -Compared VQ-VAE plus transformer decoding against an LSTM Seq2Seq baseline.
• -Explored attention mechanisms for a signal-to-language task.
• -Built a working interface around an experimental research pipeline.

Lessons

• -Signal preprocessing can dominate downstream model behavior.
• -Baselines are essential when the task itself is uncertain.
• -Experimental ML projects need careful language around what the system can and cannot infer.

What I Would Do Next

Build stronger evaluation around subject splits, noise robustness, and representation quality before expanding model complexity.