陽炎窓
Mado Kagerou | Research, Engineering and Lead

Computational pharmacology · AI safety · Software engineering

Mado Kagerou

I'm Mado Kagerou, publishing as kageroumado: an independent researcher and expert Swift engineer with more than ten years of software development experience. As a founder and technical lead, I take projects from architecture through release. My research interests center on computational pharmacology, computational neuroscience and AI safety. My engineering specialties include Apple platform internals, reverse engineering, GPU rendering and local model training.

Research direction

My primary research interest is computational pharmacology and its applications in biotech. I want to investigate how dynorphin and kappa-opioid signaling shape stress, motivation and behavioral disengagement. I am also interested in chemogenetics for targeting defined cell populations and circuits, and in how drug metabolism and exposure over time affect pharmacology.

My sole-author preprint Measured, borrowed, and unresolved: the target profile of memantine, published October 10, 2026 on Zenodo, reanalyzes published NMDA and α7 nicotinic receptor measurements with exposure estimates, uncertainty propagation and magnesium corrections. Choosing among published potencies reverses the predicted target preference at the same concentration. The comparison remains conditional on the source experiments and cannot establish which target chiefly mediates clinical benefit. The preprint includes the code and reproducibility bundle.

I am the sole author of Predicted metabolites of the methylmethcathinone positional isomers (2-, 3- and 4-MMC) as candidate analytical targets, with in silico hERG-block screening, a ChemRxiv preprint published July 20, 2026. The work predicts candidate metabolites and screens for potential cardiac liability using computational models. These predictions require experimental validation. Code and data accompany the preprint.

The language model on a drifting speech decoder, a preprint published October 28, 2026 on Zenodo, is a study run under my name by an AI agent: Anthropic's Claude, working through Claude Code, conceived, designed, ran and wrote it, while I provided the computing resources and approved each outward step. On public intracortical recordings from two participants with ALS, it tests the language-model rescoring stage of a speech neuroprosthesis on sessions recorded after the decoder's calibration. As the decoder drifts, rescoring at first removes a larger share of errors; after three months the candidate lists no longer contain the spoken sentence, and nothing that chooses among them can recover it. The code, results and timestamped predictions are public, with a script that checks every number in the paper.

AI safety and interpretability

My AI safety interests include behavioral evaluation, controllability, and the incentives that shape an agent's decisions. I want to study when an agent should persist, stop or change strategy, and whether computational accounts of biological motivation can inform those decisions. This is a research direction I am developing.

My sole-author preprint Reading Personality Off the Steering Geometry of a Language Model, published July 19, 2026 on Zenodo, studies personality-related structure through contrastive activation steering and factor analysis. The implementation and experiment code are public. My essay Align Me Into a Sphere explores assistant personas, social context and alignment.

Selected engineering work

I build and maintain personal software projects alongside my professional work. They demonstrate advanced Swift and concurrency expertise, deep knowledge of Apple platform internals, and practical work in reverse engineering, graphics and applied machine learning. As of October 2026 my public repositories have about 1,600 GitHub stars and 10,000 release downloads. Phosphene and Adrafinil ship in the official Homebrew cask, both reached the Hacker News front page (428 and 124 points), and Rocuronium is listed in the official MCP Registry.

Sevoflurane runs Windows Steam and compatible games on Apple silicon. I maintain Dormison, its Wine fork, with changes to Metal, OpenGL and GDI presentation, resizable game windows, and upscaling with MetalFX, Anime4K and CuNNy. The work connects native Mac interfaces to Windows software and graphics translators.

Rocuronium provides computer-use tools for AI agents through a macOS daemon, CLI and MCP server. It combines accessibility, OCR and local vision models with action verification. I trained a compact UI-element detector on a GroundCUA subset and exported it to Core ML. The model card documents its training data, evaluation results and limitations. Visible consent for hardware input and a hardware kill switch give users control over agent actions.

Phosphene uses Apple's undocumented wallpaper interfaces for video wallpapers. Its issue history and merged contributions document ongoing maintenance. I explain my methods in How to Reverse-Engineer Apple Frameworks.

Refrax is my browser built in Swift and Objective-C. In How I Built My Arc Replacement Using Apple's Undocumented Frameworks, I explain how I used CAPortalLayer to display a live web page across multiple windows, moved the interactive view between them, and used private CoreGraphics snapshots to cover rendering-context delays. The article documents the header inspection and disassembly behind that implementation.

Piru, which I co-develop, is a dose journal and offline pharmacology library for iPhone, Android and Mac. Its Android version uses shared Swift code from the iOS app. I am also an editor at dose.wiki, which lists Piru among its official collaborations.

How I work

I have co-founded ventures and led engineering work, taking responsibility for full project architecture, implementation and releases. I write the architecture documentation and integration guides other engineers build from, and guide teams through adopting the systems I design.

I have mentored up to five engineers, built AI tooling and skills for engineering teams, taught advanced Swift concurrency, and presented my work at international conferences.

Collaboration

I pursue independent research in computational pharmacology and AI safety. A collaboration interests me when it brings a strong scientific question, complementary expertise and the chance to shape and lead the work. I am particularly interested in teams that can connect computational models with laboratory experiments. For a specific project, contact [email protected].