Research Scientist · LLM Evaluation · AI Deployment Strategy · Production ML for Regulated Industries
Dr. rer. nat. in Genomics and AI from the Technical University of Munich and Helmholtz AI. 3 years evaluating LLMs across 28 models and three major families. Builder of clinical AI systems with EU AI Act governance. 7 peer-reviewed publications including Nature Communications.
I know where AI works, where it fails, and how to help organisations deploy it.
Within the broader GenomicsForOneHealth collection (One Health group at the University of Zurich, supervised by Lara Urban), my main contributions are in environmental metagenomics and food safety.
Links: Email · LinkedIn · ORCID
Repository layout and conventions: see STRUCTURE.md.
flowchart TB
subgraph H["Helmholtz Munich · One Health Genomics"]
direction LR
A["Aerobiome<br/><i>ISME Communications</i> 2024"]
W["Wetland multi-omics<br/><i>AEM</i> 2025"]
L["<i>Listeria</i> metagenomics<br/><i>bioRxiv</i> preprint 2026"]
end
subgraph I["Independent · LLM Evaluation"]
E["<b>Against Plausibility</b><br/>28 models · 476 step-results<br/>KW + LMM + flagship comparison"]
end
subgraph T["Applied ML · Healthcare"]
direction LR
DP["disease-progression<br/>survival + multistate"]
DN["disease-network<br/>D3.js clinical atlas"]
end
subgraph D["AI Deployment Strategy"]
direction LR
ADR["ai-deployment-readiness<br/>maturity self-assessment"]
LCC["llm-cost-calculator<br/>cost modeling + ROI"]
end
subgraph C["Consulting · AI Enablement"]
direction LR
MAD["multi-agent-dd<br/>due-diligence showcase"]
ACA["academy<br/>AI enablement curriculum"]
end
A -. ground truth .-> E
W -. ground truth .-> E
DP --- DN
E -. deployment lessons .-> ADR
E -. pricing analysis .-> LCC
ADR -. enablement .-> ACA
classDef helm fill:#e8f4fd,stroke:#1a73e8,color:#0b1b2b
classDef ind fill:#fef7e0,stroke:#f9ab00,color:#2b1b00
classDef appl fill:#e8f5e9,stroke:#34a853,color:#0b1f10
classDef depl fill:#fce8e6,stroke:#ea4335,color:#2b0b0b
classDef cons fill:#f3e8fd,stroke:#a142f4,color:#1f0b2b
class A,W,L helm
class E ind
class DP,DN appl
class ADR,LCC depl
class MAD,ACA cons
The peer-reviewed bioinformatics pipelines on the left double as ground truth for the LLM evaluation in the centre. The "plausible but wrong" failure mode that emerged from that evaluation is operationalized in the applied-ML portfolio on the right, and the deployment lessons feed into the AI strategy tools at the bottom.
My scientific work develops nanopore sequencing workflows for pathogen surveillance across air, water, food, and clinical settings. These pipelines serve as the validated ground truth for the LLM evaluation below.
| Project | Status | Repository |
|---|---|---|
| Air monitoring by nanopore sequencing | Published · ISME Communications 2024 | GenomicsForOneHealth/Air_Metagenomics · Pipeline overview |
| Wetland multi-omics surveillance | Published · AEM 2025 (shared first author) | GenomicsForOneHealth/Wetland_Health · Pipeline overview |
| Listeria metagenomic surveillance | Preprint · bioRxiv 2026 | GenomicsForOneHealth/Listeria-Adaptive-Sampling · Pipeline overview |
- Reska T, Pozdniakova S, Urban L. Air monitoring by nanopore sequencing. ISME Communications (2024).
- Perlas A*, Reska T*, Sanchez-Cano A, et al. Real-time genomic pathogen, resistance, and host range characterization from passive water sampling of wetland ecosystems. Applied and Environmental Microbiology (2025). Shared first authorship.
- Sauerborn E, Corredor NC, Reska T, et al. Detection of hidden antibiotic resistance through real-time genomics. Nature Communications (2024).
- Muchaamba F, Reska T, Biggel M, Locken KM, Weilguny L, Corti S, Kelbert L, Stephan R, Urban L. A metagenomic framework for rapid Listeria monocytogenes surveillance in food production environments. bioRxiv preprint (2026).
- Varzanadi AR, Reska T, et al. Environmental screening detects Batrachochytrium dendrobatidis. Global Ecology and Conservation (2025).
A structured benchmark of 28 models across 476 scored step-results testing whether LLMs can build real scientific workflows — not just produce locally plausible code. Two validated nanopore pipelines serve as ground truth: a 7-step aerobiome workflow and a 10-step multi-omics wetland workflow.
The dominant failure mode: "plausible but wrong" — generated code that runs, looks reasonable, and would pass a surface-level review, but makes domain-specific choices that an expert would immediately reject. These failures are invisible to automated benchmarks and dangerous in regulated environments.
No model in any family produces a fully correct 10-step wetland pipeline. Three of those steps have zero fully-correct answers across all 28 evaluated entries.
Wetland pipeline (10 steps x 28 evaluated entries). The light columns on the right are where every model in every family fails.
Flagship comparison: GPT-5 vs Opus 4.6 vs Gemini 3 Pro. Differences are not statistically significant on aerobiome — the bottleneck is benchmark statistical power, not family advantage.
Models that achieve 95%+ on isolated coding tasks can still fail 40–60% of the time on sequential pipeline tasks. Surface-level benchmarks are insufficient for deployment decisions in regulated environments. Any organisation deploying AI for multi-step analytical workflows needs domain-expert evaluation, not just code-correctness metrics.
The scoring framework (tool selection, parameter accuracy, output compatibility, scientific validity, executability) generalises to any domain where AI-generated workflows must be validated against expert ground truth — legal document drafting, financial model construction, clinical decision support, or any multi-step analytical pipeline.
Links: Full evaluation · Evaluation framework · Curated findings · Scoring criteria
Survival analysis and transformer-based architectures on synthetic FHIR/OMOP longitudinal records. Two clinical tracks (cardiovascular disease, Type 2 diabetes). Cox PH, multistate Markov, DeepSurv, DeepHit, SurvTRACE. Fairness auditing, model cards, GDPR/EU AI Act regulatory context.
Links: disease-progression/
D3.js browser dashboard for visualising disease state transitions and comparing intervention scenarios across a patient trajectory. Eight disease nodes, evidence-linked uncertainty bands, patient profile controls. No framework dependencies.
Links: disease-network/
Open-source CLI tool for organisations evaluating AI deployment. 25 questions across five dimensions (data infrastructure, process maturity, governance, talent, executive sponsorship), weighted scoring, maturity tier assignment (Exploring / Piloting / Scaling / Optimizing), sector benchmarks, and prioritised recommendations.
Built on published AI maturity frameworks (Gartner, Stanford HAI, McKinsey). 34 tests.
cd ai-deployment-readiness && python -m src.questionnaire --demoLinks: ai-deployment-readiness/
Cost modeling for LLM deployment decisions. Pricing database for 13 models across Mistral, OpenAI, Anthropic, Google, and self-hosted options. Monte Carlo sensitivity analysis (volume, token length, price variation → P10/P50/P90 projections). ROI calculator comparing manual process cost vs LLM-augmented workflow with human review layer. 23 tests.
cd llm-cost-calculator && python -m src.calculator --compare mistral-large-latest,gpt-4o,claude-sonnet-4 --input-tokens 500 --output-tokens 200 --requests-per-day 1000Links: llm-cost-calculator/
Briefing on the Claude agentic-systems frontier — primitives, topologies, ambient automation, and optimization patterns — produced by seven Haiku 4.5 subagents fanned out in parallel and synthesized by Opus 4.7. The architecture is the demonstration.
Links: agentic_systems/
Weekly digest of the LLM-coding ecosystem (Anthropic/Claude Code, OpenAI/Codex, Gemini CLI, open-source agents, IDEs, MCP, arXiv), curated to what is new since the last run. The public face of the /llm-coding-radar workflow.
Links: ai-news-radar/
Tim_Reska/
├── index.html portfolio hub
├── assets/ shared brand tokens + nav (css/tokens.css, js/nav.js)
├── llm-eval/ LLM benchmark: 28 models, 2 pipelines, 476 step-results
│ ├── methodology/ evaluation framework, scoring rubrics, pipeline references
│ ├── prompts/ standardised prompts per pipeline step
│ ├── evaluations/ scored evaluations by step and by model
│ ├── results/ figures, tables, scoring matrix
│ └── scripts/ heatmap, radar, statistical analysis generators
├── disease-progression/ survival + transformer models on FHIR/OMOP data
├── disease-network/ D3.js interactive clinical atlas
├── nanopore-advisor/ Nanopore protocol recommendation wizard
├── agent-skill-pack/ agent skill showcase
├── multi-agent-dd/ multi-agent due-diligence showcase (public teaser)
├── academy/ AI enablement curriculum overview
├── ebola/ Ebola outbreak dashboard
├── hanta/ Hantavirus outbreak dashboard (+ German de/)
├── pipelines/ pipeline overviews (→ GenomicsForOneHealth)
│ ├── aerobiome/
│ ├── wetland-surveillance/
│ └── listeria-adaptive-sampling/
├── ai-deployment-readiness/ AI maturity self-assessment (25 questions, 5 dimensions)
├── llm-cost-calculator/ LLM deployment cost + ROI modeling (13 models)
├── agentic_systems/ Claude agentic-systems frontier primer
└── ai-news-radar/ LLM-coding ecosystem weekly digest
- Reska T, Pozdniakova S, Urban L. Air monitoring by nanopore sequencing. ISME Communications (2024).
- Perlas A*, Reska T*, Sanchez-Cano A, et al. Real-time genomic pathogen, resistance, and host range characterization from passive water sampling of wetland ecosystems. Applied and Environmental Microbiology (2025). Shared first authorship.
- Sauerborn E, Corredor NC, Reska T, et al. Detection of hidden antibiotic resistance through real-time genomics. Nature Communications (2024).
- Muchaamba F, Reska T, Biggel M, et al. A metagenomic framework for rapid Listeria monocytogenes surveillance in food production environments. bioRxiv preprint (2026).
- Varzanadi AR, Reska T, et al. Environmental screening detects Batrachochytrium dendrobatidis. Global Ecology and Conservation (2025).
- Urban L, Perlas A, Francino O, et al. Real-time genomics for One Health. Molecular Systems Biology (2023).
A fuller publication record is available on LinkedIn.
- Long-read sequencing: Nanopore, PacBio Kinnex, adaptive sampling, AMR detection
- ML/AI: XGBoost, Cox PH, CTMC, transformers (SurvTRACE), SHAP, conformal prediction
- LLM evaluation: multi-vendor benchmarking, sequential pipeline testing, domain-expert scoring
- AI deployment: maturity assessment, cost modeling, ROI analysis, EU AI Act governance
- Infrastructure: Python, PyTorch, scikit-learn, lifelines, Streamlit, D3.js, Snakemake