Heidelberg University · Medical Doctoral Researcher
Clinical researcher building AI systems for biomedical research.
I develop research projects independently, from the clinical question and computational method to implementation, evaluation and usable software.
Embeddings · Knowledge Graphs · Scientific AI · Recursive Self-Improvement
Research collaborations + Postdoctoral opportunities · 2027 ↗
01 / RESEARCHER & BUILDER
Clinical Researcher · Computational Scientist · AI Systems Developer
AI METHODS & RESEARCH DIRECTIONS
Represent & retrieve
My focus is on useful representations for scientific information: semantic retrieval, similarity and connecting related evidence.
Structure & connect
I work with entities, relations and provenance to organize research knowledge and support evidence-based reasoning.
Research direction / RSI
How can a system revise its own workflows and demonstrate that the next version is better? I am interested in evaluation, feedback and controlled iteration.
The three projects above show implemented work. My RSI research interests focus on evaluating each iteration, retaining useful changes and preventing regressions.
I ask clinical questions, annoy datasets until they answer, and occasionally recruit AI agents to help.
INDEX / CHOOSE A CHAPTER
Each link opens one chapter, so you do not have to scroll through the whole notebook.
THE OPERATING SYSTEM
I tend to work on several questions in parallel. Switching between problems often helps me see connections I would miss if I stayed inside one project for too long.
Multitasking? Maybe. Parallel processing sounds more academic.
I enjoy research most when a problem requires me to learn something I did not already know.
The topic changes. The habit does not: I like finding relationships, structures and logic between things that initially look unrelated.
ONE QUESTION, MANY WINDOWS
Why can similar biological or psychological experiences become very different symptoms, behaviours and healthcare needs in different people?
RESEARCH WORLD
I study chronic pain as a heterogeneous, multidimensional human experience shaped by adversity, distress, perception, coping and functional burden.
Pain: multidimensional. Dataset: also multidimensional. My RAM: less multidimensional.
RESEARCH WORLD
Adolescent mental health unfolds across young people, parents, families, schools and pathways to help.
RESEARCH WORLD
Is the health that medicine measures always the same as the health that people experience? Interviews, scales and clinical-entry data each reveal only part of the picture.
absence of medicalization ≠ absence of suffering
RESEARCH WORLD
Can scientific work itself be redesigned? That question eventually led me to agents, reusable workflows and the Research OS.
GITHUB / PROJECT INDEX
23 repositories: research code, tools, team work, templates, archives and games. Private repositories share a non-sensitive overview; release timing depends on research, privacy and licensing.
Public · LumosNox-bamboo/DAG-Dowhy
Auditable binary treatment/outcome estimation from an explicit DAG.
Cross-fitted AIPW risk differences, cycle/backdoor checks, propensity and balance diagnostics.
Public · LumosNox-bamboo/poor-token
Use limited model resources deliberately.
Planner/executor/escalation roles and quota-aware workflow rules.
Public · LumosNox-bamboo/SCDSG-homepage
Make a Chinese-German medical community accessible online.
Bilingual static website with accessible navigation and event pages.
Public · LumosNox-bamboo/chinese-humanizer-skill
Improve natural Chinese expression while retaining facts.
Editorial skill instructions and a deterministic phrase scanner.
Public · LumosNox-bamboo/ClaimLens
Review CV claims while minimizing unnecessary uploads.
Local document extraction, structured claim review, batch handling and optional online evidence checks.
Private · LumosNox-bamboo/CARE-Meno-workflow-private
How can menopause-related experiences and care needs be studied without losing cultural and qualitative context?
Reproducible mixed-methods workflow with data normalization, available-case summaries, respondent-text coding, classification decisions and manuscript assembly. Final analysis does not include the earlier exploratory embedding/RAG branch.
Private · LumosNox-bamboo/crc-latent
How do symptoms and their relationships change over time during colorectal cancer treatment and supportive care?
Longitudinal symptom and quality-of-life analysis, regularized symptom networks and audited manuscript pipelines; related ongoing work examines sleep, physiological burden and latent structure.
Private · LumosNox-bamboo/shanghai-health-standard-2026-stage3-nsclc
Organize evidence for a clinical standards collaboration.
Documentation and source organization; software implementation not established.
Private · CEHIC-Venture-Team/cehicc-platform
Collaborative clinical platform prototype; deterministic demo data. Individual contribution not established.
Private · HeidAI2026/agent-workspace-starter
Team starter/template; individual implementation role not established.
Template or archive · retained in the repository directory.
Private · HeidAI2026/demo-repository
Team template repository; no personal scientific accomplishment inferred.
Template or archive · retained in the repository directory.
Private · HeidAI2026/LeadForge
Team scientific screening system. Attributable personal commits are architecture/communication diagrams; no sole engine ownership or biological efficacy claim.
Private · HeidAI2026/proteindesignChallenge
Collaborative computational protein-design work; individual role and experimental binding not established.
Private · kingyang728/AI4S_CNSAgents
Team executable scientific-agent workflow. Individual implementation ownership not established in inspected commit history.
Private · LumosNox-bamboo/ai4s-cns-agent-competition-archive
Historical competition artifact; counted in inventory, not as a second independent accomplishment.
Template or archive · retained in the repository directory.
Private · LumosNox-bamboo/biomedical-data-handoff-skill
Make research data handoffs more explicit and auditable.
Skill documentation for source inventories, audit outputs and workflow selection.
Private · LumosNox-bamboo/Confera
Retrieve and track accessible conference evidence without concealing missing sources.
Resumable retrieval, raw/parsed separation, HTML parsing, optional browser/OCR routes and provenance.
Private · LumosNox-bamboo/SignalLoop
Organize event signals and trace replayable decisions.
Local memory/replay workbench with API, UI and schema/citation checks.
Private · LumosNox-bamboo/tracer-method
Link survey waves while preserving conflicts and identity constraints.
Confidence-stratified graph linkage with conflict-aware components and synthetic/public benchmarks.
Public · LumosNox-bamboo/another-today
Explore small choices through a playable daily-life simulation.
Pixel-style React game prototype with browser/PWA and desktop packaging.
Public · LumosNox-bamboo/birthday-universe
Turn a birthday message into a small explorable world.
Interactive browser scenes and narrative UI.
Public · LumosNox-bamboo/de-trash-king
Explore recycling decisions through play.
Browser game logic and interactive scenes.
Public · LumosNox-bamboo/no-ac-summer
Explore a summer without air conditioning through a playful simulation.
Vanilla JavaScript state, room/heat parameters and local save system.
PROJECT DIRECTORY
A = formal output · B = code, tests and recorded outputs · C = development record · D = early idea. These levels describe evidence, not clinical effectiveness.
29 project profiles, including ongoing studies, prototypes, team work and games. Methods, contributions and current limits are shown below.
Clinical Data Science · Public
How can heterogeneous clinical observations become interpretable predictions and explicitly qualified causal estimates?
Clinical Research · Private
Why do people with similar clinical experiences develop different symptoms, burdens and care needs?
Clinical Research · Private
How can menopause-related experiences and care needs be studied without losing cultural and qualitative context?
Clinical Research · Private
How do symptoms and their relationships change over time during colorectal cancer treatment and supportive care?
Clinical Data Science · Public
Auditable binary treatment/outcome estimation from an explicit DAG.
Clinical Research · Private
Organize evidence for a clinical standards collaboration.
Clinical Data Science · Private
Link survey waves while preserving conflicts and identity constraints.
Clinical Data Science · Private
Quantify spatial pain information from drawings.
Clinical Research · Public
Measure pain sensitivity, social-emotional impact and acceptance.
Scientific AI Systems · Private
How can a computational result be traced to its inputs, validation checks and permitted scientific claims?
Scientific AI Systems · Private
How can scientific tasks combine tools, explicit constraints, candidate generation and evaluation?
Scientific AI Systems · Private
Can protocol-bound tests distinguish history-dependent behavior from simplistic explanations?
Scientific AI Systems · Private
Explore computational candidate binders and pH-related hypotheses.
Research Engineering · Private
Make research data handoffs more explicit and auditable.
Research Engineering · Public
Improve natural Chinese expression while retaining facts.
Research Engineering · Public
Review CV claims while minimizing unnecessary uploads.
Research Engineering · Private
Retrieve and track accessible conference evidence without concealing missing sources.
Research Engineering · Public
Use limited model resources deliberately.
Software Engineering · Public
Make a Chinese-German medical community accessible online.
Software Engineering · Private
Organize event signals and trace replayable decisions.
Research Engineering · Private
Index accessible public poster evidence while retaining source identity.
Collaboration Inventory · Private
Collaborative clinical platform prototype; deterministic demo data. Individual contribution not established.
Collaboration Inventory · Private
Team scientific screening system. Attributable personal commits are architecture/communication diagrams; no sole engine ownership or biological efficacy claim.
Collaboration Inventory · Private
Collaborative computational protein-design work; individual role and experimental binding not established.
Collaboration Inventory · Private
Team executable scientific-agent workflow. Individual implementation ownership not established in inspected commit history.
Creative Coding · Public
Explore small choices through a playable daily-life simulation.
Creative Coding · Public
Turn a birthday message into a small explorable world.
Creative Coding · Public
Explore recycling decisions through play.
Creative Coding · Public
Explore a summer without air conditioning through a playful simulation.
PUBLICATIONS
Published studies, with their research questions, methods, findings and limitations.
BMC Medicine · 2026 · Published
Author names and order follow the published record.
DOI: 10.1186/s12916-026-04851-7 ↗Humanities and Social Sciences Communications · 2026 · Published
Author names and order follow the published record.
DOI: 10.1057/s41599-026-07796-7 ↗Pain Research and Management · 2026 · Published
Author names and order follow the published record.
DOI: 10.1155/prm/3924115 ↗BMC Public Health · 2025 · Published
Author names and order follow the published record.
DOI: 10.1186/s12889-025-24354-z ↗Behavioral Sciences · 2026 · Published
Author names and order follow the published record.
DOI: 10.3390/bs16081442 ↗BMJ Open · 2022 · Published
Author names and order follow the published record.
DOI: 10.1136/bmjopen-2022-061922 ↗SKETCHBOOK PAGES
Research has occasionally taken me places. European railways have occasionally taken me somewhere else.
I love travelling. European railways have occasionally interpreted this as a collaborative decision-making process. Several trips ended somewhere I did not intend to visit. I generally visited it anyway.
A VERY SHORT LAB CAREER
During my first cell-biology practical, I somehow managed to kill the entire class’s cells.
I subsequently developed a healthy respect for people who can keep cells alive.
PUZZLE CORNER
Here is a puzzle instead.
1/4Three sealed notebooks are labelled Pain, AI and Both. Every label is wrong. You may inspect one page. Which notebook do you open?
Best ways to start a conversation: an interesting dataset · a strange clinical question · a puzzle I should not spend the next hour solving
NOTES FROM THE WORKSHOP
If I have renamed the same files twice, a script is probably next.
File naming, formatting checks and literature tables take time. I build small tools to make those steps repeatable and leave a record of how an analysis was done. That leaves more time to think about the result—and occasionally discover that I asked the wrong question.
Which research tasks can we automate, and where does a researcher need to make the decision?
RESEARCH OS
The comic sketches the workshop I would like to build. Current prototypes handle local tabular analysis, validation checks, evidence tracing and rules for what a result can support. The v0.4 runtime uses deterministic code; LLM calls and manuscript generation are outside its scope.
HUMAN JUDGMENT Researchers make the final call.
METHODS IN PRACTICE
A conceptual workflow. Project notes describe the implemented modules and their current limitations.
METHODS & TOOLS
Methods I use, tools I have built, and what I am learning next. Each entry links to the relevant work.
RESEARCH SNAPSHOT
SEM · Psychometrics · Network Analysis
DoWhy · DAGs · observational causal inference
scikit-learn · SHAP · interpretable ML · predictive modeling
versioned plans · source tracking · validation gates · human review · audit logs
R · Python · Git · tests
DO NOT OPEN
2027 / POSTDOCTORAL RESEARCH & COLLABORATION
Questions I would like to pursue next.
I am completing my medical doctoral research at Heidelberg University and looking for postdoctoral opportunities in 2027. I expect to defend in spring 2027; the dates of the defense and degree award remain to be confirmed.
I would like to join an established research team, work with large clinical datasets and develop new questions with colleagues across disciplines.
01 / RESEARCH INTEREST
I want to study how research agents use tools, handle medical data and retrieve evidence, and how researchers can evaluate their outputs, detect errors and trace the supporting sources.
Relevant work includes Research OS validation checks, Confera source records and my architecture contributions within the AI4Science team.
Research OS: the current prototype →02 / RESEARCH INTEREST
I am interested in combining clinical records, psychological measures, longitudinal observations and image features to study disease heterogeneity and outcomes. Missing data, model interpretation and performance in new populations are central questions.
Published clinical methods →03 / RESEARCH INTEREST
I want to develop tools that clinical research teams can use and inspect: clear analysis steps, reproducible results and explicit points for researchers to review uncertainty.
Evidence auditing with ClaimLens →What I ultimately want to build
Medical care draws on specialist knowledge, clinical experience and the systems that help people use them. In places with fewer resources, these approaches need to fit local staff, infrastructure and patients’ lives.
How can clinical knowledge and decision support travel well, with local teams shaping their use?
My long-term dream is to help make good medicine less dependent on postcode.
Especially in psychiatry.01Which clinical questions become answerable when we can follow large cohorts over time?
02Can AI agents help scientists generate and falsify hypotheses systematically?
03How do psychosocial experiences become measurable biological or clinical burden?
04Why do medical systems see some symptoms clearly and barely see others?
05Can we connect pain, mental health, neuroinflammation and brain health using multimodal longitudinal data?
06What happens when medicine, causal inference and Scientific AI are designed together from the beginning?
07
Probably enough projects.
COLLABORATION
I like to begin with a clinical question, then work out the study design and methods together.
I am also interested in teams with strong clinical data that are beginning to build their AI / computational research capability.
ESPECIALLY INTERESTED IN
Have data? Let’s build AI together.
Agent: ‘I found one more.’Mengmeng: ‘Email me.’