What This Umbrella Record Covers
At DataJoint, I worked on the infrastructure layer for neuroscience data systems: schema-backed workflows, DataJoint Elements, cloud-oriented processing patterns, deployment support, notebooks, visualization/testing surfaces, and operational debugging across real lab data.
This page is an umbrella record. The individual deployed-system records on this site describe specific contexts such as Sabatini Lab and Allen Institute Mindscope more carefully.
The Core Problem
Modern neuroscience data is multimodal and operationally messy. A single research program can involve:
- Electrophysiology
- Calcium imaging
- Fiber photometry
- Behavioral video
- DeepLabCut pose estimation
- Facemap-style facial movement analysis
- Experimental metadata
- Analysis notebooks and review tools
The hard part is not only processing each modality. The hard part is preserving the relationships between raw acquisition, processing parameters, derived artifacts, quality control, and the scientific question.
Systems I Worked Around
DataJoint Elements
I worked in the ecosystem of reusable DataJoint workflow packages, including patterns for electrophysiology, imaging, behavior, pose estimation, and lab/session organization.
The recurring engineering theme was modularity: build workflows that can be adapted to a lab's instruments and conventions while still preserving provenance and reproducibility.
element-facemap Training and Inference Schemas
I integrated deep-learning training and inference components into element-facemap, helping make Facemap-style facial behavior analysis reproducible inside DataJoint's schema framework.
That work involved representing model configuration, training metadata, inference inputs, inference outputs, and downstream behavioral features as structured workflow components rather than loose scripts. The goal was to make facial inference deployable, testable, and reusable across lab settings.
element-deeplabcut Model Training and Inference
I added and tested model training and inference support for element-deeplabcut, extending DataJoint's pose-estimation workflows beyond simple result storage.
The important engineering work was connecting DeepLabCut project/model state, video inputs, training runs, inference jobs, and output artifacts into a workflow that could be validated and deployed. I also worked through cloud deployment issues so the deep-learning components could run in practical GPU-backed environments rather than only on a developer laptop.
Electrophysiology Workflows
Electrophysiology workflows required careful handling of large binary files, metadata, spike sorting configuration, quality metrics, waveform extraction, and downstream review.
This is where infrastructure details become scientific details: bad file discovery, unclear metadata, or untracked parameters can change what a researcher believes about a neural signal.
SpikeInterface into element-array-ephys
I worked on integrating SpikeInterface-style processing into element-array-ephys, connecting spike sorting workflows to structured DataJoint schemas.
This work helped bridge high-density electrophysiology, sorter configuration, processing parameter sets, computed result tables, and provenance. In practice, it meant making spike sorting workflows more reproducible, queryable, and deployable across local and cloud execution contexts.
Behavioral Video and Pose Workflows
DeepLabCut and Facemap-style systems connect video, trained models, inference jobs, confidence scores, and downstream behavioral features. These workflows need to be reproducible, but also flexible enough for lab-specific cameras, body parts, ROIs, and experimental paradigms.
Cloud and Deployment Work
The deployment side involved the practical realities of scientific computing:
- Containerized environments
- Database configuration
- Object storage patterns
- GPU and CPU compute jobs
- Local testing and remote execution
- Failure recovery and documentation
That work made me think of scientific infrastructure less as "a pipeline" and more as an operating environment for research.
Representative Deployed Records
Sabatini Lab DataJoint Workflow
The public Sabatini Lab workflow repo documents a DataJoint setup with Imaging, Photometry, Behavior, Ephys, and DLC behavior video organization, plus notebooks, docs, tests, and Dockerized SciViz testing. My record for that system is written conservatively around what the public repo supports and my DataJoint operations role.
Allen Institute Mindscope Operations
The Allen Mindscope record captures DataJoint workflow support for large-scale neuroscience infrastructure. Some details are not public, so the page focuses on the type of work: schema-backed scientific workflows, reproducible processing, cloud-oriented execution, and reviewable outputs.
Lu Lab Deep Learning Behavior Pipelines
The Lu Lab record captures applied deployment work for deep-learning behavioral analysis: DataJoint-backed DeepLabCut pose estimation, Facemap facial inference, cloud execution, testing, and optimization of training/inference workflows for real lab video data.
What This Shows About My Work
This work is the professional version of a pattern that appears across my portfolio: I like making complex signals usable.
In this case, the signals are scientific: files from acquisition rigs, video streams, neural traces, model outputs, and human annotations. The software job is to preserve enough structure that researchers can move from raw data to interpretation without losing trust.
That is the kind of system work I enjoy: practical, high-context, scientifically consequential, and built at the boundary between code and reality.