Jev Pokémon Trainer
An AI can know the rules and still make a poor move. I wanted a way to investigate which information was missing and whether changing the context would improve its decisions.
A battle agent, a library of game mechanics, and a decision laboratory for inspecting the context behind a move.
TypeScript · Pokémon Showdown · TypeSafe Jev
- Input
- Battle observations, legal actions, and relevant game knowledge.
- Output
- An action selection and a record of the context behind it.
- Components
- Mechanics library, context retrieval, scenario evaluation, and a Pokémon Showdown integration.
Play a battle, then explore the context, alternatives, and simulations behind Jev’s decisions. Limited public pilot: one battle at a time.
OpenCoil
A coil drawing does not tell us how its fields interact with anatomy. I wanted a workbench where I could change a design or its placement, inspect the calculated fields, and return to the settings behind each result.
A web interface and scientific backend for configuring coil simulations, inspecting fields, and preserving reproducible runs.
Python · React · TypeScript · SimNIBS
- Input
- Coil geometry, anatomical model, placement, and stimulation parameters.
- Output
- Simulated fields, orthogonal views, and saved run settings.
- Components
- Coil configuration, SimNIBS workflows, field visualization, and run comparison.
Research software.
Mayna
Coding agents become more useful when they can work with the context of a whole project. Mayna explores what that experience could look like for physicians working with a patient’s evolving history.
An AI agent workspace that brings longitudinal patient context, clinical instructions, documents, and evidence into the same workflow for physician review.
Python · FastAPI · React · OpenAI Agents SDK
- Input
- Clinical instructions, case context, and supported audio, documents, or images.
- Output
- Editable clinical drafts, contextual summaries, and evidence for review.
- Components
- Patient workspaces, agent orchestration, longitudinal context, and document workflows.
Clinical outputs remain subject to physician review.
Melanoma CNN
Could a neural network learn visual patterns useful for distinguishing skin-lesion images in a melanoma classification task? This experiment turned that question into a training and evaluation workflow.
A convolutional neural network built around a pretrained VGG19 model, with scripts for organizing images, training the classifier, and inspecting its predictions.
Python · TensorFlow / Keras · VGG19
- Input
- Labeled skin-lesion images.
- Output
- Binary classification scores and training and evaluation plots.
- Components
- Image preparation, augmentation, model training, checkpoints, and evaluation scripts.
An image-classification experiment, not a validated diagnostic tool.
Postoperative pain analysis
Why do some patients still have pain when leaving the post-anesthesia care unit? Our case-control study examined factors associated with pain at discharge.
Analysis scripts accompanying our published study, making the statistical approach available for inspection.
Python · Logistic regression
- Use
- Inspect the analysis accompanying the published study.
- Reference
- Postoperative Pain at Discharge From the Post-anesthesia Care Unit: A Case-Control Study (2024).
