AI for Insect Monitoring

Role

I first got involved through my senior design project. Early on, the sponsor told me he was impressed by how quickly I was making progress. Later in the semester, he pulled me aside and asked me to continue working with the research team after graduating.

I now work part-time as a Research Assistant. My work includes adding features to the field pipeline, training and evaluating models, building tools to monitor and debug field runs, and developing the downstream system used to organize research data after collection.

Making the pipeline observable

When I joined the project, debug information was split across temporary popups, terminal output, and generated files. That made it difficult to tell how much work was left or whether the device was still making progress.

I proposed and built a modular interface that combines the pipeline's live state and debug information in one place. The UI shows processing progress visually through video overlays, along with runtime metrics and state information that make it much easier to understand what the system is doing.

In the field, a monitor can be plugged into the device and the interface appears automatically. This makes it possible to inspect the system live instead of extracting and searching through output files.

Model work

I train and evaluate the computer vision models used in the field pipeline. I expanded the system to support finer-grained taxonomic identification.

I also made the pipeline more configurable so different field setups and processing stages can be changed without rewriting the overall workflow.

I am intentionally not publishing the current models, pipeline implimentation, or internal device details while the related research is still in progress.

Research data workflow

I proposed the initial design for a larger post-processing system and have written most of the current implementation. It takes in field data from different types of field devices, preserves metadata needed to trace each collection, and converts the outputs into a consistant format for further processing and storage on AWS.

The goal is to make large field datasets easier for researchers to work with. Instead of manually searching through individual device outputs, researchers will be able to filter the collected data and extract only the subset relevant to a particular analysis.

Publication status

A research manuscript related to this work is currently in preparation. Until that work is public, I am intentionally keeping unpublished model details, internal system information, and research results off this site.