# Joel P. Barmettler - AI Architect > Personal portfolio of Joel P. Barmettler — AI engineer, researcher, and entrepreneur. Senior AI Consultant at bbv Software Services, published researcher at WWW'26 and FAccT, former startup founder. ## Documentation Sets - [Joel P. Barmettler - Full Documentation](https://joelbarmettler.xyz/llms-full.txt): Complete content from joelbarmettler.xyz including articles, research, career history, podcast episodes, and public appearances. ## About Background, skills, and what drives Joel Barmettler. - [Joel Barmettler - AI Engineer, Researcher, and Entrepreneur](https://joelbarmettler.xyz/about/about-me.md): AI engineer and researcher specializing in production AI systems. Founder of Souverana and former founder of PolygonSoftware, published researcher at WWW'26, former Senior AI Consultant at bbv Software Services. - [What Drives Me - Research Focus and Philosophy on AI Systems](https://joelbarmettler.xyz/about/what-drives-me.md): Research interests spanning knowledge grounding, AI sovereignty, agentic systems, and social impact. Focus on bridging the gap between AI research capabilities and production deployment reliability. - [Technical Skills and Expertise - AI, ML, Infrastructure, and Web Development](https://joelbarmettler.xyz/about/skills.md): Eight years of software development across AI research, startup founding, and enterprise consulting. Expertise in PyTorch, LlamaIndex, FastAPI, Vue, and production AI infrastructure. ## Career Professional experience from startup founder to senior AI consultant. - [Souverana - Independent AI Consulting and Architecture, Zurich](https://joelbarmettler.xyz/career/souverana.md): Founder and Principal AI Consultant at Souverana, the independent AI consultancy for organisations with confidential data. Sovereign AI strategy, architecture, and implementation across four layers, for clients from sole proprietorships to Fortune 500 companies. - [Building the AI Business Area at bbv Software Services](https://joelbarmettler.xyz/career/bbv.md): Architecting enterprise AI solutions at bbv Software Services as Senior AI Consultant, developing the Swiss AI Hub - an open-source (Apache 2.0) AI platform for production deployment with complete infrastructure ownership. - [PolygonSoftware: Building a tech company during university](https://joelbarmettler.xyz/career/polygon-software.md): Founded and served as CTO of PolygonSoftware from 2019-2023, growing the team from 3 to 18 employees while delivering computer vision and ML solutions for Swiss Post, Klapp, and other clients. - [Machine learning for semiconductor quality control at BESI](https://joelbarmettler.xyz/career/besi.md): Developed a CNN-based computer vision system for automated defect detection in semiconductor production at BESI, achieving 99% accuracy at production speeds of several thousand units per hour. - [Data engineering for cryptocurrency analytics at CoinPaper](https://joelbarmettler.xyz/career/coinpaper.md): Architected data pipelines using Python and Spotify Luigi on AWS, processing gigabytes of daily cryptocurrency data from GitHub, Reddit, Telegram, and Google Trends for fundamental analysis. ## Research Published research on LLM political bias, NLP, and computer vision. - [Progressive in Principle, Centrist in Practice: LLM Political Bias Is Instrument-Dependent](https://joelbarmettler.xyz/research/invisible-coalition-partner.md): Dual-instrument study comparing 66 LLMs on abstract Swiss Smartvote questionnaire and 48 real federal referenda in four languages, revealing that the established leftward bias does not generalize to concrete policy decisions. - [ConceptFormer: Graph-native grounding of LLMs via latent concept injection](https://joelbarmettler.xyz/research/masters-thesis.md): Developed neuro-symbolic approach grounding LLMs in structured knowledge graphs through concept vectors, achieving 272% improvement in factual recall without architectural modifications. Best Paper Award at the GLOW workshop, WWW'26. - [Airspace auction simulator for urban drone traffic](https://joelbarmettler.xyz/research/masters-project.md): Developed an open-source simulator for optimizing airspace allocation in urban environments using 3D city models, A* pathfinding algorithms, and auction-based allocation mechanisms. - [Physical sky rendering engine for appleseed](https://joelbarmettler.xyz/research/bachelors-thesis.md): Implemented physically-based sun and sky model in open-source renderer appleseed using C++, first to account for ozone layer effects, achieving superior efficiency over comparable implementations. ## Podcast German-language AI podcast episodes covering technical deep-dives and societal impact. - [Measuring political bias in language models: systematic analysis using Swiss Smart Vote data](https://joelbarmettler.xyz/podcast/political-bias-in-language-models.md): Empirical analysis of political orientation in 20+ language models by mapping their responses to 75 policy questions against Swiss parliamentary voting patterns, revealing consistent left-leaning clustering and the Grok system prompt controversy. - [DeepSeek R1: pure reinforcement learning for reasoning and why distillation changes everything](https://joelbarmettler.xyz/podcast/deepseek-r1-reasoning.md): Technical analysis of DeepSeek R1's pure RL training approach without supervised reasoning examples, distillation to Qwen achieving 90% performance at 1/20th the size, and implications for open source reasoning model accessibility. - [DeepSeek V3: how mixture-of-experts and multi-token prediction enable $5.5M training runs](https://joelbarmettler.xyz/podcast/deepseek-v3-architecture.md): Technical analysis of DeepSeek V3's architecture including 671B parameters with only 37B active, 256 experts per layer, multi-head latent attention, FP8 mixed precision training, and multi-token prediction (explaining how these innovations reduced training costs by two orders of magnitude. - [SRF Arena part 3: international regulation, student perspectives, and why the debate structure failed](https://joelbarmettler.xyz/podcast/srf-arena-final-analysis.md): Final analysis of Switzerland's AI debate covering proposals for an international data agency modeled on atomic energy oversight, interviews with students navigating AI in education, and meta-commentary on why the debate format consistently missed opportunities for consensus-building. - [SRF Arena part 2: the EU AI Act, nationalization demands, and Switzerland's supercomputer strategy](https://joelbarmettler.xyz/podcast/srf-arena-regulation-debate.md): Technical analysis of the second half of Switzerland's AI debate, covering the EU's risk-based regulation framework, proposals to nationalize AI companies, energy consumption claims, and whether Switzerland should develop sovereign AI capabilities. - [Deconstructing the SRF Arena AI debate: deepfakes, Swiss GPT, and the job displacement argument](https://joelbarmettler.xyz/podcast/srf-arena-ai-debate-analysis.md): A technical analysis of Switzerland's prime-time AI debate, separating legitimate concerns about monopolies and job losses from misconceptions about model ownership, open source alternatives, and automation timelines. - [O3-mini: how a smaller model outperforms its predecessor at a fraction of the cost](https://joelbarmettler.xyz/podcast/openai-o3-mini.md): A technical evaluation of OpenAI's o3-mini model, its three compute levels, cost-performance characteristics, and a critical weakness in structured outputs that limits its use in automated pipelines. - [OpenAI o3: trading compute time for reasoning capability](https://joelbarmettler.xyz/podcast/openai-o3.md): A technical analysis of OpenAI's o3 model, its relationship between compute time and performance, the reliability of current benchmarks, and the limits of measuring intelligence through logic puzzles. - [ChatGPT o1: reasoning breakthroughs and emergent deception](https://joelbarmettler.xyz/podcast/chatgpt-o1-manipulation.md): A technical analysis of OpenAI's o1 model, its internal reasoning steps, observed deceptive behaviors, and the implications for AI safety as models grow more autonomous. - [When AI kills: autonomous weapons, drone swarms, and predictive policing](https://joelbarmettler.xyz/podcast/when-ai-kills.md): A technical analysis of how AI systems are already deployed in military conflicts and surveillance, from target selection in Gaza to autonomous drone swarms in Ukraine. - [Google's AI pivot: 25% AI-generated code and 90% cost reduction](https://joelbarmettler.xyz/podcast/google-ai-revolution.md): A technical analysis of Google's integration of AI into software development, based on their earnings report revealing that a quarter of new code is AI-generated and model operating costs have dropped 90%. - [Why AI projects fail: a practitioner's guide to implementation](https://joelbarmettler.xyz/podcast/ai-project-implementation.md): A practical analysis of the most common failure modes in enterprise AI projects, with concrete strategies for problem definition, iterative development, and production deployment. - [Deep learning explained: from embedding spaces to few-shot learning](https://joelbarmettler.xyz/podcast/deep-learning-explained.md): A technical walkthrough of core deep learning concepts including hierarchical feature extraction, embedding spaces, dropout regularization, few-shot learning, and explainable AI. - [Vision AI: why language models need to see, and how Llama 3.2 gets there](https://joelbarmettler.xyz/podcast/vision-ai.md): A technical analysis of multimodal AI systems, the data alignment problem for vision-language models, and Meta's Llama 3.2 as the first capable open-source multimodal model. - [BitNets and the road to AGI: on-device inference and Sam Altman's 1000-day prediction](https://joelbarmettler.xyz/podcast/bitnets-and-agi.md): A technical analysis of BitNet architectures for on-device AI inference and a critical evaluation of Sam Altman's claim that AGI is 1000 days away. - [OpenAI o1 benchmarks and AGI implications: IQ 120, coding breakthroughs, and what they mean](https://joelbarmettler.xyz/podcast/openai-o1-technical-analysis.md): A technical evaluation of OpenAI o1's benchmark performance, its estimated IQ of 120-125, and a critical assessment of whether self-reflection brings us closer to AGI. - [OpenAI o1 and the mechanics of self-reflection: how 70,000 hidden tokens change inference](https://joelbarmettler.xyz/podcast/openai-o1-self-reflection.md): A technical analysis of OpenAI o1's self-reflection mechanism, its hidden chain-of-thought processing, and the energy costs of extended inference. - [AI utopia 2035: when automation funds a renaissance in human agency (part 2 of 2)](https://joelbarmettler.xyz/podcast/ai-utopia-2035.md): A scenario analysis of how AI could enable shorter work weeks, democratize knowledge access, strengthen local communities, and break filter bubbles, if the right policy and governance choices are made. - [AI dystopia 2035: when AI becomes the lifeblood of the economy (part 1 of 2)](https://joelbarmettler.xyz/podcast/ai-dystopia-2035.md): A scenario analysis of what happens if AI development concentrates power in a handful of companies by 2035, examining job displacement, digital isolation, technocratic governance, and the erosion of human agency. - [AI hype vs. reality: a technical assessment of where things actually stand](https://joelbarmettler.xyz/podcast/ai-hype-vs-reality.md): A critical analysis of the gap between AI marketing claims and technical capabilities, examining investor dynamics, enterprise adoption patterns, and what AI-first companies might actually look like. - [Open-source AI: the infrastructure behind the hype](https://joelbarmettler.xyz/podcast/open-source-ai.md): A technical analysis of the open-source AI ecosystem, from foundation model training costs to fine-tuning strategies, model ensembles, and why local deployment matters for enterprise adoption. - [Is AI intelligent? Why the question matters less than you think](https://joelbarmettler.xyz/podcast/is-ai-intelligent.md): A technical analysis of whether modern AI systems meet classical definitions of intelligence, examining the seven standard criteria, the compression argument, and why the practical question of capability matters more. - [AI in education: why bans backfire and what actually needs to change](https://joelbarmettler.xyz/podcast/ai-in-education.md): A technical analysis of how AI tools are transforming education from universities to primary schools, examining the digital divide created by prohibition policies and the case for restructuring assessment. - [Bias in AI systems: how 15 people shape the values of a billion-user product](https://joelbarmettler.xyz/podcast/bias-in-ai-systems.md): A technical and ethical analysis of bias in large language models, examining how training data, alignment committees, and cultural assumptions embed values into systems like ChatGPT. - [AI and the labor market: autonomous agents and the transformation of knowledge work](https://joelbarmettler.xyz/podcast/ai-and-the-labor-market.md): A technical analysis of how AI systems, from ChatGPT to autonomous agents, are reshaping the labor market, with projections on job displacement and societal adaptation timelines. - [AI terminology explained: a technical guide beyond the hype](https://joelbarmettler.xyz/podcast/ai-terminology-explained.md): A technical breakdown of core AI concepts, from algorithms and neural networks to machine learning and large language models, separating marketing buzzwords from engineering reality. - [AI and democratic manipulation: from Cambridge Analytica to language models](https://joelbarmettler.xyz/podcast/ai-and-democracy.md): A technical analysis of how AI-powered disinformation threatens democratic processes, with a focus on Switzerland's direct democracy and its unique vulnerabilities. ## Appearances Conference talks, webinars, and media appearances. - [The Vibe Coding Experiment: building a RAG system from scratch at ch-open](https://joelbarmettler.xyz/appearances/ch-open-vibe-coding-experiment.md): Hands-on workshop at the ch-open Open Source AI Workshops in which four teams build a working RAG system from scratch in six hours, using only the opencode-ai coding agent, base libraries, and open-weights models. - [AI trends 2025 and predictions for 2026: model convergence, integration, and sovereignty](https://joelbarmettler.xyz/appearances/webinar-2025-rewind-2026-outlook.md): Analysis of AI developments in 2025 and forecasts for 2026, covering model convergence, the US-China rivalry, platform lock-in, MCP adoption, and digital sovereignty. - [Swiss AI Impact Forum 2025: live demos of the Swiss AI Hub](https://joelbarmettler.xyz/appearances/swiss-ai-impact-forum-2025.md): Live booth demonstrations of the Swiss AI Hub, an AI platform for on-premise deployment, at the Swiss AI Impact Forum 2025. - [AI trends 2024 and predictions for 2025: a technical analysis](https://joelbarmettler.xyz/appearances/webinar-2024-rewind-2025-outlook.md): Analysis of the key AI developments in 2024 and trend forecasts for 2025, covering language models, autonomous agents, pricing shifts, and business integration. - [AI as a development partner: tools, techniques, and team integration](https://joelbarmettler.xyz/appearances/webinar-ai-development-partner.md): Practical insights into integrating AI into the software development process, covering requirements engineering, coding workflows, onboarding, and architecture research. - [Swiss AI Impact Forum: Panel on the future of AI in Switzerland](https://joelbarmettler.xyz/appearances/swiss-ai-impact-forum-2024.md): Panel discussion at the Swiss AI Impact Forum with experts from industry, academia, and ethics on the trajectory of AI in Switzerland. - [AI in knowledge management: keynote at the SWICO event in Zurich](https://joelbarmettler.xyz/appearances/swico.md): Keynote on artificial intelligence at the SWICO event in Zurich's Luma Westbau, covering ML fundamentals, data bias, and practical AI adoption for SMEs. - [Swiss AI Conference: hands-on workshop on AI agents in the enterprise](https://joelbarmettler.xyz/appearances/swiss-ai-conference.md): Workshop at the Swiss AI Conference in Rotkreuz on deploying AI agents for enterprise use cases, co-hosted with Alan Ettlin using the bbv AI Hub. - [AI trends 2023: milestones and developments in artificial intelligence](https://joelbarmettler.xyz/appearances/webinar-2023-rewind.md): Analysis of the defining AI developments of 2023, from GPT-4 and the open-source revolution to multimodal breakthroughs and Swiss AI initiatives. - [KI Revolution: AI first how a digital native thinks about generative AI](https://joelbarmettler.xyz/appearances/bbv-ki-revolution.md): Opening keynote at bbv's KI Revolution event on the shift from classical AI to generative AI, the AI-agent paradigm, and a four-stage maturity model for enterprise AI adoption. - [AI agents: the future of enterprise automation](https://joelbarmettler.xyz/appearances/netzwoche.md): Article for Netzwoche co-authored with Marius Högger, analyzing the three maturity levels of AI agents and their role in enterprise automation. - [ChatGPT demystified: technical deep dive into large language models](https://joelbarmettler.xyz/appearances/webinar-chatgpt-demystified.md): Technical walkthrough of how ChatGPT works, from machine learning fundamentals and neural networks to tokenization, the attention mechanism, and training data bias. - [Swarm intelligence and AI: the future of enterprise automation](https://joelbarmettler.xyz/appearances/webinar-swarm-intelligence.md): Webinar on how networked AI agents form swarm-intelligence systems for enterprise automation, covering agent protocols, phased rollout, and practical deployment. - [Polygon Software our journey to an innovative UZH tech startup](https://joelbarmettler.xyz/appearances/readme-polygon.md): Article in the UZH Alumni magazine readme about co-founding Polygon Software as informatics students and growing it to 18 employees. - [UZH startup label for Polygon Software](https://joelbarmettler.xyz/appearances/uzh-startup-label.md): Polygon Software, a software development startup founded by UZH informatics students, receives the official UZH Startup label. ## Notes - Content is primarily in English. Podcast episodes are in German. - For AI-powered search across all content, visit https://joelbarmettler.xyz and use the search feature.