Problem: Segmenting semantically meaningful text from engineering drawings (such as blueprints and schematics) is a rather more difficult job than, for example, OCR on text from standardized printed pages like books, newspapers, or magazines.
Approach: Fine-tuning the Kraken convolutional neural network, which was trained on pages or images with more distortion and other abberations. I developed and ran ML experiments for optical character recognition, iterating between data preprocessing and experimental testing. As part of the process, I streamlined and improved the data processing pipeline, improved training data accounting and quality, and improved results assessment and validation. Finally, I contributed to the final report for the Department of Energy grant supporting this research.
A series of data science tutorials and open source evangelism for Onboard Data, Inc, including:
- tutorials on the company's API and API client [one, two, three]
- timeseries cleaning and basic imputation techniques [colab, medium]
- feature engineering and selection [colab, medium]
- timeseries forecasting with Facebook's Prophet [colab, medium]
- outlier and anomaly detection [colab, medium]
- fault detection in HVAC systems [medium]
A Bayesian Model of Cue-based Cardinal Direction Estimation [pdf]
Absolute Direction Feedback Impacts Environmental Knowledge [pdf]
How Do You Know If You’re Lost or Not? Epistemic and Pragmatic Action During Navigation [pdf]
Navigational Feedback Technology Alters Environment Awareness [pdf]
Lost and Confused: Measuring Uncertainty in Navigation [pdf]
- Developing company's R API client from scratch following CRAN standards.
- Refined the company's internal R client, bringing it up to CRAN standards, making it ready to be public-facing. Improved parity with company's public-facing Python API client. Updated the company's API client docs (ReadTheDocs) to reflect the newly-public R client.
- Overhauled one portion of this open source AHU fault detection. Reduced redundant code, shifted to OOP, improved modularization.