{"person":{"name":"Carl Luis Pöhl","pronouns":"they/them","photo":"/carl.jpg","roles":["Clinical AI","LLM Evaluation & Safety","Oncology Decision Support","AI Consulting"],"location":"Munich · open to relocation","availability":"Available January 2027","email":"clpoehl@gmail.com","links":{"linkedin":"https://www.linkedin.com/in/carl-luis-poehl","github":"https://github.com/cl-poehl","orcid":"https://orcid.org/0000-0002-6358-1581"},"profile":"Final-year medical student, computer-science graduate and PhD candidate specialising in clinical AI and oncology decision support. Co-founder of AI consultancy Halverd, with experience developing, evaluating and deploying research and commercial AI systems. Secured four university hospitals for a multi-site clinical AI study. Final medical examination expected December 2026. Four first-author manuscripts complete, in co-author review. Poster accepted for the ESMO AI & Digital Oncology Congress 2026."},"sections":[{"id":"ventures","label":"Ventures","entries":[{"id":"halverd","title":"Halverd","org":"AI consultancy for professional services","url":"https://halverd.com","role":"Co-founder & AI Consultant","period":"06/2026 – present","bullets":["Shipped multiple AI systems into production, spanning candidate matching, team planning and assistants inside the messaging channels people already use, each returning output with the reasoning attached.","Built the agent infrastructure for a client deployment and the evaluation framework behind it, with test sets and scoring so output quality stays measurable after release.","Co-delivered a £15,000+ recruitment-sector build, integrated with the workflow software their consultants already use."],"tags":["Agents","Evals","Production"]},{"id":"concordaince","title":"ConcordAInce","org":"Tumor-board decision-support AI","url":"https://concordaince.com","role":"Co-founder","period":"12/2025 – present","bullets":["Translating the PhD tumor-board decision-support work toward clinical deployment, with privacy-preserving infrastructure that keeps patient data within the hospital perimeter.","Dresden Exists LifeTech incubator. Y Combinator Spring 2026 interview."],"tags":["Oncology","On-prem","Startup"]},{"id":"healicus","title":"Healicus","org":"Natural-medicine AI assistant","url":"https://healicus.com","role":"Co-founder","period":"03/2026 – present","bullets":["Consumer-facing RAG reference for natural medicine, with each entry grounded in a Cochrane review, EMA HMPC monograph, EFSA authorised health claim or major-journal RCT.","Personalised pre-generation checks for drug interactions, contraindications and allergies, with severity-tiered alerts and auditable event records."],"tags":["RAG","Safety","Consumer"]}]},{"id":"research","label":"Research","entries":[{"id":"phd","title":"Multi-agent decision support for the prostate-cancer tumor board","org":"Dept. of Urology, University Hospital Dresden","role":"PhD Candidate (Dr. rer. medic.), AI in Medicine","period":"01/2025 – present","summary":"Supervised by Sherif Mehralivand and Marcus Sondermann, in cooperation with Jakob Kather and Lars Hilgers (Kather Lab, EKFZ).","bullets":["Developed hybrid multi-agent tumor-board decision support combining structured patient data with EAU and German S3 oncology guidelines.","Retrospective validation at the prostate-cancer tumor board: 96.6% guideline adherence, with additional guideline-compliant therapy options identified in 57.8% of cases. External validation across four university hospitals, including a prospective arm, ongoing.","Introduced guideline conformity as a separate evaluation endpoint after finding that the fine-tuned system agreed with tumor-board decisions in 99.2% of cases yet ranked below every structured design on conformance.","Compared 44 configurations × 500 held-out real cases across six architectures. Validated the LLM judge against eight board-certified raters, each scoring the same 50 cases, and designed their scoring rubric and adjudication protocol. Multi-agent, tree and hybrid designs were most guideline-conformant on intermediate- and high-risk cases, while zero-shot, RAG and fine-tuned approaches performed better on easy cases.","Study leadership: secured participation from Heidelberg, Mannheim, Zurich and Basel for the multi-site study, persuading two initially sceptical full professors without formal authority. Wrote every ethics submission, reworking it through two rounds of committee feedback to approval. Secured the project's HPC compute allocation on a written proposal.","Architecture: combined verified LLM extraction, deterministic decision trees, multi-agent escalation, case retrieval and PubMed evidence. The published benchmark evaluates an earlier configuration of this system.","Data: benchmarked on 1,000 tumor-board cases drawn at random from an 11,488-case institutional archive, split 500 test and 500 development, with fine-tuning on 2,521 separate archive cases and zero patient-level overlap. Held-out cases were excluded from retrieval, and retrieval was never combined with fine-tuned models.","Deployment: integrated the open-source system into the hospital infrastructure, running on-premise in the clinic on commodity hardware so no patient data leaves the hospital. GDPR-compliant by design.","Safety: built automated evaluation and CI/CD checks for guideline adherence, evidence traceability, hallucinations and contraindications, with cite-or-abstain behaviour and clinician oversight. Red-teamed clinical failure modes and built regression tests using constitutional-AI-style synthetic cases."],"tags":["Multi-agent LLMs","LLM-as-judge","Clinical validation"]},{"id":"pancreas","title":"Pancreatic histogenomics","org":"DKFZ Heidelberg + University Hospital Heidelberg, Surgery","period":"12/2025 – present","summary":"In collaboration with Andrea Bauer.","bullets":["Benchmarked seven foundation-model encoders on 527 fresh-frozen pancreatic whole-slide images, achieving 96.8% balanced accuracy on the 475-slide PDAC/pancreatitis/normal subset in patient-level cross-validation.","Evaluated calibrated abstention on the five-class differential: 95.1% accuracy at 77% coverage (confidence ≥0.90). External PDAC-vs-rest detection achieved 0.936 AUROC on 257 TCGA-PAAD frozen slides.","Evaluated pathology-transcriptomics fusion on 303 slides from 301 patients. Rare cystic and neuroendocrine tumours received greater genomic weighting across four random seeds.","Developed an inflammatory-infiltrate model on 311 pancreatic slides: 83.2% balanced accuracy and 0.897 macro-AUROC in patient-grouped cross-validation. Its predictions tracked nine immune-gene readouts more closely on average than pathologist grades (mean Spearman ρ 0.535 vs. 0.462 across 203 matched slides)."],"tags":["Pathology FMs","Calibration","Multimodal"]},{"id":"mdthesis","title":"Calibrated Parkinson’s progression-subtype prediction","org":"Dept. of Neurology, University Hospital Dresden","role":"MD Thesis (Dr. med.)","period":"04/2025 – present","summary":"Supervised by Tom Hähnel.","bullets":["Developed calibrated Parkinson’s progression-subtype prediction from 17 routine scores in 409 PPMI patients, using per-patient OLS slopes and intercepts.","Compared Random Forest, XGBoost and logistic regression with missing-score handling and class-conditional conformal abstention.","Assessed fixed-model predictions against motor, cognitive and biomarker outcomes in the later PPMI 2.0 cohort. Independent subtype-level validation remains outstanding."],"tags":["Conformal prediction","Neurology"]},{"id":"katherlab","title":"DNA language models across 19 cancer types","org":"Kather Lab (EKFZ / TU Dresden)","role":"Student AI Researcher","period":"04/2024 – 11/2024","bullets":["Benchmarked Hyena-DNA, DNABERT-2 and NucleotideTransformer-v2 across 19 TCGA cancer types (7,674 patients). Hyena-DNA embeddings with a residual feed-forward network achieved 0.94 mean-class AUROC at 54.96% overall accuracy.","Improved NucleotideTransformer-v2 accuracy by almost 10% with mutation-centred windows, approaching DNABERT-2. Results suggested greater sequence-shift sensitivity with k-mer than byte-pair tokenisation."],"tags":["Genomic FMs"]}]},{"id":"opensource","label":"Open source","entries":[{"id":"cite-or-abstain","title":"cite-or-abstain","org":"Python · MIT","url":"https://github.com/cl-poehl/cite-or-abstain","summary":"Clinical LLM evaluation harness for citation verification, unsupported confidence and abstention, with human-validated LLM judging, DeepEval integration and reproducible audit reports.","tags":["Evals","DeepEval"]},{"id":"promotionshub","title":"promotionshub","org":"TypeScript · Next.js · live","url":"https://promotionshub.de","summary":"Platform where medical students find doctoral positions and read verified reviews of supervision, live with 135 positions across 28 research groups. Ratings are structurally separated from paid listings so promotion cannot influence them, named scores stay hidden until four independent verified reviews corroborate them, and verification documents are discarded after checking.","tags":["Product","Trust & safety"]},{"id":"parkinson-subtype-predictor","title":"parkinson-subtype-predictor","org":"Python · MIT · live demo","url":"https://github.com/cl-poehl/parkinson-subtype-predictor","summary":"Live Streamlit app with calibrated predictions, abstention, SHAP and counterfactual explanations. Single-patient and batch workflows, plus reproducible PPMI analysis respecting the data-use agreement. Demo: parkinson-subtype-predictor.onrender.com.","tags":["Streamlit","Explainability"]},{"id":"hf-titration-assistant","title":"hf-titration-assistant","org":"FastAPI + React/TypeScript · MIT","url":"https://github.com/cl-poehl/hf-titration-assistant","summary":"Heart-failure prototype combining explainable deterioration risk with guideline-based GDMT titration. XGBoost/LightGBM on 98 trajectory features. Zigong evaluation with patient-level splits and bootstrap confidence intervals.","tags":["Cardiology","XGBoost"]}]},{"id":"writing","label":"Manuscripts & talks","entries":[{"id":"manuscripts","title":"Manuscripts in preparation","summary":"First-author throughout. Four manuscripts complete, in co-author review.","bullets":["PROMPT: a deterministic guideline tree achieves similar automated guideline-conformance scores to multi-agent LLMs for prostate cancer tumour-board decisions. Pöhl, Sondermann, Kather, Mehralivand.","Pan-pancreatic diagnosis on fresh-frozen sections: a foundation-model characterization of the open multi-class differential, calibrated abstention, and class-conditional multimodality. Pöhl et al. DKFZ Heidelberg + University Hospital Heidelberg.","A foundation model reads a molecularly anchored inflammatory-infiltrate axis from fresh-frozen pancreatic histology. Pöhl et al. University Hospital Heidelberg + DKFZ Heidelberg + University of Verona.","A Calibrated, Abstaining Prediction Model for Parkinson’s Disease Progression Subtypes from a Variable Set of Routine Clinical Scores. Pöhl, Falkenburger, Hähnel. TU Dresden. Basis of the Dr. med. thesis."]},{"id":"talks","title":"Selected talks","bullets":["Nov 2026: Poster (accepted), ESMO AI & Digital Oncology Congress, Berlin: multi-agent decision support for the prostate-cancer tumor board.","Sep 2026: Invited lab meeting, Dewey Lab, University of Cambridge: “Agreeing with the tumor board is not the same as being right.”","Jul 2026: Invited talk, DKFZ Heidelberg: prostate-cancer decision support.","Jun 2026: Invited talk, Urology, University Hospital Basel: guideline-grounded AI.","May 2026: Invited lab meeting, Krauthammer Lab, University of Zurich: clinical LLMs.","Mar 2026: Poster, ELSA TrustworthyAI4Health: prostate-cancer tumor boards.","Oct 2024: Invited talk, Connectome Fall Symposium: DNA-language models."]}]},{"id":"earlier","label":"Industry","entries":[{"id":"ey","title":"EY, Munich","org":"Transaction, Strategy & Execution","role":"Intern","period":"08/2022 – 09/2022","bullets":["Co-developed the kick-off strategy for a client's market exit and prepared execution workshops with global stakeholders."]},{"id":"bayer","title":"Bayer, Tokyo","org":"Crop Science Division","role":"Data Science Intern","period":"02/2020 – 03/2020","bullets":["Built the division's first data-driven targeting model for sales-activity planning, with no in-house data science function to build on, replacing gut-feel prioritisation. A retrospective backtest indicated savings of up to ¥120M. Presented findings to the head of the Asia-Pacific division."]}]},{"id":"education","label":"Education","entries":[{"id":"doctorates","title":"Dr. rer. medic. + Dr. med.","org":"TU Dresden","period":"2025 – expected Q1 2027","bullets":["Dr. rer. medic. (PhD equivalent), AI in Medicine: multi-agent decision support for the prostate-cancer tumor board.","Dr. med. (medical doctorate, research thesis): calibrated progression-subtype prediction in Parkinson’s disease."]},{"id":"medicine","title":"Medical studies (State Examination)","org":"TU Dresden · Semmelweis University, Budapest","period":"09/2020 – present","bullets":["Second national medical examination (M2) passed 2025, good (2.0), ~top 20% nationally. Final examination December 2026.","At TU Dresden since 10/2022, after pre-clinical studies at Semmelweis University, Budapest (09/2020–07/2022). Physiology and Biochemistry examinations at the top grade.","International clinical rotations across six settings: cardiology (Buenos Aires) · surgery + dermatology (Windhoek, Namibia) · visceral surgery + neurosurgery (University Hospital Heidelberg) · gastroenterology + haematology/oncology (University Hospital Zurich) · radiology (TUM University Hospital, Munich) · general practice (southern Germany)."]},{"id":"computerscience","title":"B.Sc. Computer Science","org":"RWTH Aachen · EPFL · HKUST","period":"2017 – 2024","bullets":["RWTH Aachen (10/2018–11/2024): final grade good (2.0), completed part-time alongside medicine. Mathematics very good (1.3), minor in Business Administration.","EPFL Lausanne (09/2019–03/2020): exchange semester in Computer Science, SEMP scholarship.","HKUST Hong Kong (09/2017–10/2018): Dual Degree Program in Technology & Management (Computer Science & Business). GPA 3.74/4.0, Dean's List both semesters.","Abitur, German School Tokyo Yokohama (05/2017): final grade 1.6, Mathematics and Physics as written subjects. Youngest graduate, at age 16.","Grading scales: German 1.0 best, 4.0 pass · HKUST 4.0 best."]},{"id":"awards","title":"Awards, scholarships & programmes","bullets":["McKinsey Capstone Programme, member (2020 – present).","Dresden Exists LifeTech Incubator Program, incubator place for ConcordAInce (2025).","Porsche IT Scholarship, competitive technology scholarship (2020–2021).","BASF Science Prize (Chemistry) · German Physical Society Prize (Physics), Abitur 2017.","SEMP Scholarship, Swiss-European Mobility Programme, EPFL exchange (2019–2020).","HKUST Dean's List, top 10% of cohort (2017–2018)."]},{"id":"teaching","title":"Teaching & academic engagement","bullets":["Tutor, Biochemistry, Faculty of Medicine, TU Dresden (04/2023–06/2025): review and exam-preparation sessions for cohorts of 100+ medical students.","Executive Committee Member, Connectome Neuroscience Society, TU Dresden (10/2023–10/2025): organised interdisciplinary neuroscience and neurosurgery seminars and workshops."]}]},{"id":"toolkit","label":"Toolkit","entries":[{"id":"skills","title":"Technical skills","bullets":["Core ML: PyTorch · NumPy · pandas · scikit-learn · XGBoost · statsmodels","LLMs & training: LoRA / QLoRA · multi-agent orchestration · prompt engineering · structured outputs · Anthropic / OpenAI SDKs · vLLM","Evaluation & safety: Inspect · DeepEval · OpenAI Evals adapters · LLM-as-judge · red-teaming · regression testing · evidence attribution · calibration · prompt-injection mitigation","Biological foundation models: GigaPath · UNI / UNI2 · Virchow2 · CONCH · CTransPath · TITAN · ABMIL / CLAM-MB / TransMIL / DTFD-MIL · gated cross-modal fusion · genomic sequence models","Retrieval & RAG: FAISS · Chroma · pgvector · LangChain · LlamaIndex · hybrid retrieval","Backend & deployment: FastAPI · Next.js · PostgreSQL · Supabase · AWS · Hetzner · Docker · CI/CD · GitHub Actions · Streamlit · pytest · Git · on-premise deployment"]},{"id":"languages","title":"Languages & interests","bullets":["German, Spanish and English (native) · French (fluent) · Italian (conversational) · Japanese and Mandarin (basic).","Raised across Ecuador, Mexico, Spain, Germany, Japan and Norway.","Concert piano performance · chess · alpine skiing · sailing."]}]}]}