Work, sorted by how it was judged.

How I knew it worked

Class imbalance was the central risk: a model that calls everything “pass” scores well and is useless. I used stratified splits and SMOTE, and judged the model on its ability to catch the failures rather than on overall accuracy.

• Python
• pandas
• regex
• scikit-learn
• XGBoost
• LightGBM

How I know it's working

The whole design is bounded by a <50 ms latency budget for real-time inference on embedded hardware, so accuracy that arrives late is not accuracy. The evaluation set is built deliberately adversarial: sirens at distance, sirens under traffic noise, and sounds that resemble sirens but are not.

• Real-time audio ML
• Embedded / low-latency inference
• Crowdsourced acoustic data pipeline

How I knew it worked

The TF-IDF feedforward model won at 93.19% accuracy and 0.9296 macro F1, against 92.17% and 0.9205 for the LSTM. The metric choice was the real decision.

• Python
• TensorFlow/Keras
• scikit-learn
• NLTK
• TF-IDF
• LSTM

How I knew it worked

71.9% multiclass accuracy across 12 species, 100% on the binary task, and 78–85% confidence on held-out field recordings the model had never seen. More useful than the headline number was the error analysis: the confusion matrix showed the model mixing up House Finch and White-crowned Sparrow specifically, which traced to genuinely similar call structures rather than to a modelling bug.

• Python
• TensorFlow
• CNNs
• librosa
• STFT / spectrograms

How I knew it worked

Compared Linear Regression against Random Forest and reached R²=0.86, meaning the model accounts for 86% of demand variance. The finding worth having was behavioural: occupancy runs 25% higher on rainy days (r=0.78).

• Python
• pandas
• Linear Regression
• Random Forest
• Seattle Open Data API
• Open-Meteo API

How I knew it worked

The correctness bar here is not a metric, it is whether the times match what the masjid actually posted, which is why the two-tap community correction path exists at all. Design constraint held throughout: zero tracking, so the app cannot be validated by watching users, only by making the correction loop cheap enough that people use it..

• Swift
• SwiftUI
• Native iOS
• Lock Screen widget
• Location services
• Apple Maps