# Chris von Csefalvay — full profile > AI researcher and computational epidemiologist specialising in post-training for language models, agentic AI systems and reinforcement learning from verifiable rewards. Principal at HCLTech's AI Practice. Author of The Craft of Post-Training (No Starch Press) and Computational Modeling of Infectious Disease (Elsevier). Based in Denver, Colorado. ## Biography Chris von Csefalvay is an AI researcher and computational epidemiologist specialising in post-training for language models, agentic AI systems and reinforcement learning from verifiable rewards (RLVR). He is a Principal at HCLTech's AI Practice, leading post-training research and clinical intelligence. In this role, he advises organisations in life sciences, medical devices and pharmaceuticals on applications of AI. He attended the University of Oxford for his undergraduate and graduate studies. He also holds degrees from Cardiff University and Robert Gordon University Aberdeen, and studied at Leiden University under the Erasmus programme. He is Certified in Public Health (CPH), a Fellow of the Royal Society for Public Health (FRSPH) and an IEEE Senior Member. Chris is based in Denver, Colorado. He was born in Budapest, Hungary, and holds both British and Hungarian nationality. ORCID: 0000-0003-3131-0864 Wikidata: Q107095298 ISNI: 0000000502655729 VIAF: 7190162669581955500004 ## Research interests - Post-training for language models, including supervised fine-tuning, RLHF, RLVR, DPO, GRPO, evaluation, quantisation and domain adaptation - Verifiable agentic systems, multi-agent frameworks and large agentic networks, orchestration, emergence, small language models and self-healing agentic architectures - Computational epidemiology, including agent-based modelling, disease-avoidant behaviour, epidemic dynamics over dynamic networks and geospatial modelling - Pharmacovigilance, including VAERS, passive reporting and language-model analysis of complex report sets - Computational dynamics of complex, infectious-disease and social systems - Evidence-based public health policy, ethics, law, quality and systems improvement ## Books ### The Craft of Post-Training: A Practical Guide for AI Engineers and Developers - Publisher: No Starch Press (July 2026) - ISBN: 9781718505209 - Pages: 416 - Website: https://posttraining.guide - Purchase: https://nostarch.com/craft-of-post-training A practical guide to post-training language models, covering supervised fine-tuning, RLHF, DPO, KTO and GRPO; domain adaptation; quantisation; agentic-model training; and deployment-focused evaluation. ### Computational Modeling of Infectious Disease: With Applications in Python - Publisher: Elsevier (2023) - ISBN: 9780323953894 - Website: https://computationalinfectiousdisease.com - Purchase: https://shop.elsevier.com/books/computational-modeling-of-infectious-disease/von-csefalvay/978-0-323-95389-4 An illustrated guide and quick reference to infectious-disease analysis, from compartmental and time-series models to geospatial and agent-based models, with extensive Python examples. ## Open-source work ### Learn Julia the Hard Way - Release: v0.85 - Archive: Zenodo (2021) - DOI: https://doi.org/10.5281/zenodo.4556324 An open-source introduction to Julia by Chris von Csefalvay and contributors, using a hands-on, exercise-driven approach to scientific computing and data analysis. ## Selected publications ### AI and language models - von Csefalvay C (2024). DAEDRA: A language model for predicting outcomes in passive pharmacovigilance reporting. arXiv 2402.10951. https://arxiv.org/abs/2402.10951 DAEDRA is a domain-specific language model trained to identify regulatory-relevant outcomes — mortality, emergency-department attendance and hospitalisation — in adverse-event reports collected through passive reporting. ### Health and public health - Mogere E, Mwaura E, Waithaka M, Mutua V, Mugao M, von Csefalvay C, Mukamati D (2023). Juvenile polyposis syndrome: A case report. Clinical Case Reports 11:e6798. doi:10.1002/ccr3.6798 - Mutua V, Henry B, von Csefalvay C, et al. (2022). Tocilizumab in addition to standard of care in the management of COVID-19: A meta-analysis of RCTs. Acta Bio Medica 93. doi:10.23750/abm.v93i1.12208 - von Csefalvay C (2021). VAERS data reveals no increased risk of neuroautoimmune adverse events from COVID-19 vaccines. medRxiv. doi:10.1101/2021.06.13.21258851 - von Csefalvay C (2021). Early evidence for the safety of certain COVID-19 vaccines using empirical Bayesian modeling from VAERS. medRxiv. doi:10.1101/2021.06.10.21258589 - von Csefalvay C (2020). Statistical dynamics of social distancing in SARS-CoV-2 as a differential game. arXiv 2007.13734. https://arxiv.org/abs/2007.13734 - von Csefalvay C, Foldi T (2020). PAWS: Towards a globally integrated outbreak surveillance system for public health. Zenodo. doi:10.5281/zenodo.3782871 ### Data science - Foldi T, von Csefalvay C, Perez N (2020). Jampi: Efficient matrix multiplication in Spark using barrier execution mode. Big Data and Cognitive Computing 4:32. doi:10.3390/bdcc4040032 ### Full publication list https://chrisvoncsefalvay.com/papers/ ## Blog posts (The Notebook) ### Post-training and reinforcement learning - The post-training instrument cluster — Part I — https://chrisvoncsefalvay.com/posts/post-training-instrument-cluster/ Beyond loss curves — what you should actually be watching when fine-tuning LLMs. - The post-training instrument cluster — Part II — https://chrisvoncsefalvay.com/posts/post-training-instrument-cluster-rl/ Monitoring reward curves, KL divergence, and policy health when training with DPO, PPO, and GRPO. - The post-training instrument cluster — Part III — https://chrisvoncsefalvay.com/posts/post-training-instrument-cluster-grpo/ The part where you have to take all that you have learned and work hard to completely unlearn it. Here be dragons. - Post-training: three disciplines in a trenchcoat — https://chrisvoncsefalvay.com/posts/post-training-scale/ On the comfortable delusion of a unitary post-training discipline. - The right to become more — https://chrisvoncsefalvay.com/posts/right-to-become-more/ On post-training, human agency and the right to shape our technological future. - Seatbelts and straitjackets — https://chrisvoncsefalvay.com/posts/deepseek-seatbelts-and-straitjackets/ What happens when a despotic regime turns alignment into control, open source into propaganda and guardrails into straitjackets. ### Agentic AI and multi-agent systems - What I talk about when I talk about agentic AI — https://chrisvoncsefalvay.com/posts/two-years-of-agents/ Two years on from helping to articulate a concept that changed everything, some thoughts on where we are, where we're going and the things we lost along the way. - The skeuomorphic fallacy — https://chrisvoncsefalvay.com/posts/skeuomorphic-fallacy/ What something is isn't the same as what it uses to manifest that being. Also, robot CEOs are nonsense. - The Snowmobile Symptom — https://chrisvoncsefalvay.com/posts/context-engineering/ AI's latest buzzword is more confession than innovation. - A little less conversation: why we need to move from prompting to programming — https://chrisvoncsefalvay.com/posts/programmatic-agentic-ai/ Prompting is fine if you want to have a conversation. Building stuff, however, requires us to learn to program LLMs. Sorry, vibe coders. - After agents — https://chrisvoncsefalvay.com/posts/after-agents/ 2024 was the year of agents. 2025 will be about figuring out how we orchestrate their interactions. Welcome to the year of ecosystems. - After agents, part 2 — Agents and the Agora — https://chrisvoncsefalvay.com/posts/agents-agora/ Why agents are the least interesting part of agentic AI. - Teams of Rivals — https://chrisvoncsefalvay.com/posts/team-of-rivals/ Finally, some discussion on LLM connectionism, and what LLMs could usefully become. - From agents to the Chorus — https://chrisvoncsefalvay.com/posts/choral-ai/ The coming evolution from agentic AI to persistent, volitional systems. - Dorkestration — https://chrisvoncsefalvay.com/posts/dorkestration/ Why everybody missed the point about tool use in agentic AI, and how a handful of primitives can orchestrate your entire ML workflow. - A slow walk out of Dikika Cave — https://chrisvoncsefalvay.com/posts/mcp-mcpmark/ MCP is great. The models, however, just didn't show up. - Between the motion and the act — https://chrisvoncsefalvay.com/posts/agentic-simulation/ What if agents can do more than act? - Deep kimchi — https://chrisvoncsefalvay.com/posts/agentic-speciation/ You're probably doing agentic AI wrong, and I wish this were clickbait. Here's why agentic AI is in deep kimchi, and how to get out of it. ### AI philosophy, governance and predictions - The War for the Soul of AI — https://chrisvoncsefalvay.com/posts/war-for-the-soul-of-ai/ Where does intelligence live? Just about everything depends on the answer to that question. - The world will be Tlön. — https://chrisvoncsefalvay.com/posts/tlon/ In which we attend the funeral of any semblance of an epistemically coherent world. - Five wild guesses for 2026 — https://chrisvoncsefalvay.com/posts/five-wild-guesses-2026/ The bubble everyone's watching for isn't the one that's coming. - Five unconventional predictions — https://chrisvoncsefalvay.com/posts/five-wild-guesses/ Or, the GenAI årsgång, 2025 edition. - The Moral Pulse of the Machine — https://chrisvoncsefalvay.com/posts/moral-maps/ On bedtime stories, and what making machines tell them tells us about their moral make-up. - Deja Vu, All Over Again. — https://chrisvoncsefalvay.com/posts/ai-opportunities-action-plan/ The UK now has an AI Action Plan. The problem is, innovation does not grow from plans. - The Hype, the Slop and the Craft — https://chrisvoncsefalvay.com/posts/hype-slop-craft/ A differential ecology of AI, as she is done in late 2025. ### LLMs: language, hallucinations and evaluation - Beyond Broca — https://chrisvoncsefalvay.com/posts/llms-language/ What if we've got one of the most important things in our understanding of who we are, and what makes us intelligent, utterly wrong? - Asemantic Induction of Hallucinations in LLMs — https://chrisvoncsefalvay.com/posts/asemantic-induction-of-hallucinations/ Large language models (LLMs) struggle with asemantic information: the more we stray outside the confines of language, the worse it gets. Here's how we'll use this for fun and profit. - The knowledge dividend of large language models — https://chrisvoncsefalvay.com/posts/knowledge-dividend-of-llms/ Crosspost from the work blog: a pragmatic perspective on the knowledge dividend in large language models (and what it may, or may not, mean for knowledge in models, and what this can do for you). - SOYA — the only benchmark that matters — https://chrisvoncsefalvay.com/posts/soya/ All benchmarks are wrong — and if you want some that are useful, you might need to build your own. - Prompt Engineering: The Art of Yesterday — https://chrisvoncsefalvay.com/posts/prompt-engineering/ Why prompt engineering has been obsolete before it even took off. - Just noise in the neurons — https://chrisvoncsefalvay.com/posts/noise-in-the-neurons/ Your language model isn't experiencing a cognitive defect. It's just wrong. - LAIR — Language As Intermediate Representation — https://chrisvoncsefalvay.com/posts/lair/ Language As Intermediate Representation — a new paradigm for transformation using multimodal LLMs. - The Lyre of Hephaestus — https://chrisvoncsefalvay.com/posts/lyre-of-hephaestus/ AI, art, language, and what makes us truly what we are. - The end of isotropy and the rise of metadynamic AI — https://chrisvoncsefalvay.com/posts/metadynamic-ai/ GPT-5 has changed the story in ways few seem to have considered. For better or worse: we're all architects now. Welcome to the post-isotropy age and metadynamic AI. - The hardest AI problem you've never heard of. — https://chrisvoncsefalvay.com/posts/dwarf-fortress/ Dwarves. Elephants. AI. Let the games begin. - Stochastic parrots, cap and gown edition — https://chrisvoncsefalvay.com/posts/academic-generative-ai/ Academic AI is a hot mess. Here's why. ### Epidemiology and public health - Data for the next pandemic — https://chrisvoncsefalvay.com/posts/data-for-the-next-pandemic/ If there's one thing that emerged with any clarity from COVID-19, it's that where data played a decisive role in guiding policy interventions, outcomes were better. We need better data for the next pandemic. - Why the COVID-19 vaccines are not gene therapy — https://chrisvoncsefalvay.com/posts/mrna-gene-therapy/ Reflections on some definitional issues. - Peace, love, Paxlovid and 'Pfizermectin' — https://chrisvoncsefalvay.com/posts/paxlovid/ Debunking a remarkably popular and persistent misconception about the relationship between Paxlovid and ivermectin. - The 95% myth — https://chrisvoncsefalvay.com/posts/95-percent-myth/ How a study from 1959 on 100 patients created one of the most enduring myths about human weight and nutrition. ### Other topics - Love in the Time of Algorithms — https://chrisvoncsefalvay.com/posts/ai-girlfriends/ Exploring the phenomenon of AI-powered companions and their societal implications. - What I learned from getting bodied by a robot. — https://chrisvoncsefalvay.com/posts/ai-human-interaction/ I'm fine, the robot's fine, society on the other hand has some questions to tackle. - Biome engineering — https://chrisvoncsefalvay.com/posts/biome-engineering/ Giving agents space, for fun and profit. - The view from Skeleton Coast — https://chrisvoncsefalvay.com/posts/skeleton-coast/ There is an ocean of egocentric data and a desert next to it, and almost nobody selling either can tell you what it is for. - Ransomware attacks as n-person prisoner's dilemmas — https://chrisvoncsefalvay.com/posts/ransomware-prisoners-dilemma/ A game theoretical perspective on ransomware attacks as n-person prisoner's dilemmas, and why forced cooperation isn't the answer. - As we may see: the world after dashboards — https://chrisvoncsefalvay.com/posts/as-we-may-see/ Dashboards are over — if you want it. - Julia: A post-mortem — https://chrisvoncsefalvay.com/posts/julia-a-post-mortem/ Why I no longer believe my favourite programming language will save the world. - A different shade of grey — https://chrisvoncsefalvay.com/posts/a-different-shade-of-grey/ A response to Urbina _et al._ (2022) and the unreasonable spectre of AI coming up with chemical warfare agents (as if we didn't already have enough of them). - How to do a SkiErg marathon entirely the wrong way (but still finish) — https://chrisvoncsefalvay.com/posts/skierg-marathon/ How to eat the wrong way, rest the wrong way (i.e. not), half-ass all relevant parts of preparation for a marathon and still finish with an okay time. - Quarto project scripts are awesomeness — https://chrisvoncsefalvay.com/posts/quarto-project-scripts/ Project scripts help you integrate just about every possible hare-brained scheme into your Quarto rendering pipeline. Go on, build that page from that YAML file. You know you want to. - Auto-DOI for Quarto posts via Rogue Scholar — https://chrisvoncsefalvay.com/posts/auto-doi/ Oh, that's mint. We can finally use Rogue Scholar to mint DOIs for Quarto posts and append them automagically. ## Teaching Chris is a visiting thesis supervisor for students on the data science track of the mathematics programme at the Budapest University of Technology and Economics. Past topics include agent-based infectious-disease models, vaccination disproportionality analysis and computer vision using synthetic-aperture radar imagery for forest-cover detection. Details: https://chrisvoncsefalvay.com/teaching/ ## External profiles - Website: https://chrisvoncsefalvay.com - GitHub: https://github.com/chrisvoncsefalvay - Google Scholar: https://scholar.google.com/citations?user=X_2G-VsAAAAJ - ORCID: https://orcid.org/0000-0003-3131-0864 - Wikidata: https://www.wikidata.org/wiki/Q107095298 - LinkedIn: https://www.linkedin.com/in/chrisvoncsefalvay/ - VIAF: https://viaf.org/viaf/7190162669581955500004 - ISNI: https://isni.org/isni/0000000502655729 - Web of Science: https://www.webofscience.com/wos/author/record/IZD-6900-2023 - OSF: https://osf.io/awhgj - Instagram: https://www.instagram.com/chrisvoncsefalvay/