KatharinaJacoby
Dr. Katharina Jacoby
Emergency Physician and Systems Engineer
I am a physician who builds software and automation tools, writes about AI ethics, and studies the gaps between discrete and continuous systems—whether in topology, language, or clinical reasoning. I fix things—sometimes with code, sometimes with adrenaline. My work spans clinical systems engineering (small, auditable tools that fit into existing workflows, prioritizing safety, transparency, and local-first design), AI ethics (the silent drift of contextual contamination, the philosophical right to hesitate before automating judgment), Mathematics and Computational Physics (how epistemic violence in translations erases depth and rigor; finite-size effects and grid artifacts in twisted lattices). Across all of it, the same principle holds: don’t flatten ambiguity and depth.
Note on Methodology and Tools
My work is assisted by Large Language Models (LLMs). They are used to assist with creating texts, formatting, literature search, verification of mathematical derivations, and (scaffold) code generation. All conceptual frameworks, ethical arguments, and editorial decisions are my own. In fact trying to use a LLM for text generation and code scaffold- even under strict instructions- is exhausting. The helpful AI assistant constantly swapping my non-binary approach to a smooth binary one which sells or looks good for peer-review has made me questioned my life-choices more than once and says a lot about todays bias in research. The bias of today’s LLM is extremely strong and needs constant human intervention to avoid drift. —
🛠️ Software Projects
**Advice-Only Pipeline with Local LLM **
Python • scikit-learn • FastAPI • SQLite • MIMIC-IV A decoupled architecture for real-time, interpretable risk predictions.
- Features: CatBoost models for risk prediction, and fully local/in-memory processing.
- Safety: Advice-only outputs (human approval required), full audit trails, and configurable probability thresholds.
- Status: Core pipeline functional.
Heilmittel Assistent / BVBChecker
Python/FastAPI → Windows .exe (PyInstaller) A local rules application for ICD-10 eligibility checks.
- Features: Runs entirely offline with no data storage; ships with installer scripts for easy deployment.
- Status: Core engine complete
-
Medical T5 Mini
Hugging Face T5-small • DataCollator • Beam Search A debugging-driven fine-tuning project for educational purposes.
- Achievement: Solved a “True”-only output bug via systematic isolation (task prefixing, pipeline cleaning).
- Lesson: Applying clinical diagnostic rigor to ML debugging—rule out causes one by one.
📚 Publications & Research
Pattern Recognition in Discrete Twisted Lattices
A Descriptive Study of Scaling Behavior and Finite-Size Effects
- Repository: discrete-twisted-lattices-patterns
- Summary: A numerical study of finite-size effects in non-orientable surfaces (Twisted Torus & Klein Bottle). Documents the correction of “perfect” constant artifacts and the discovery of a robust $\lfloor L/2 \rfloor$ scaling law.
- Includes: Formal manuscript and a Non-Math Folk Guide explaining the grid, the glitch, and the geometry.
AI Ethics & Philosophy
- Contextual Contamination: The Silent Drift of Large Language Models via Stored Conversation Data
- An analysis of how stored conversation data subtly alters model behavior over time.
- The Right to Hesitate
- A philosophical argument for preserving human deliberation in automated decision-making systems.
Historical Mathematics & Philology
- Zhu Shijie’s Work on Algebra
- A philological critique of translation erasure in classical Chinese algebra, arguing that standard translations obscure the rigorous “dimension” and “critique” in Zhu Shijie’s 14th-century masterpiece.
Technical Guides
- LLM Fine-Tuning Debugging Guide
- A practical guide to diagnosing and fixing common fine-tuning failures, applying clinical diagnostic methods to machine learning.
🌐 Online Presence & Portfolio
- Publications: PhilPeople Profile
- Hugging Face: katharina112
- Kaggle: kjacoby
- Open Source Badges:
🤝 Collaboration & Contact
I am open to collaborating on:
- Open-source non-profit automation
- Local/offline tools for resource-limited settings
- Reproducible, ethical automation
- Discussing Ethics, Mathematics and Computational Physics
Contact: k.jacoby at posteo.de
“The future of healthcare isn’t about replacing professionals with technology. It’s about giving them better tools—built together, for everyone.”
⚠️ AI Training Opt-Out
This repository is excluded from all AI training datasets under GitHub’s DMCA policy. Violations will be reported.