talk-data.com

LLM-based web agents have recently made significant progress, but much of it has occurred in closed-source systems, widening the gap with open-source alternatives. Progress has been held back by two key challenges: first, a narrow focus on single-step tasks that overlooks the complexity of multi-step web interactions; and second, the high compute costs required to post-train LLM-based web agents.

To address this, we present the first statistically grounded study on compute allocation for LLM web-agent post-training. Our approach uses a two-stage pipeline, training a Llama 3.1 8B or QWEN 2.5 7B student to imitate a Llama 3.3 70B teacher or QWEN 2.5 72B via supervised fine-tuning (SFT), followed by on-policy reinforcement learning (GRPO).

We find this process highly sensitive to hyperparameter choices, making exhaustive sweeps impractical. To spare others from expensive trial-and-error, we sample 1,370 configurations and use bootstrapping to estimate effective hyperparameters. Our results show that combining SFT with on-policy RL consistently outperforms either approach alone on both WorkArena and MiniWob++. Further, this strategy requires only 55% of the compute to match the peak performance of pure SFT on MiniWob++, effectively pushing the compute-performance Pareto frontier, and is the only strategy that can close the gap with closed-source models.

Read the paper on ArXiv: How to Train Your LLM Web Agent: A Statistical Diagnosis (PDF)

About the speaker

I’m Massimo Caccia, Senior Research Scientist at ServiceNow Research, specializing in post-training methods for computer-use agents. I see computer use as the ultimate playground for testing agents, thanks to its ubiquity and diversity. My research involves conducting large-scale empirical studies to systematically evaluate trade-offs among different approaches and to develop practical know-how, with reinforcement learning being a particular focus.

As a core contributor to the web-agent research library ecosystem, I actively shape evaluation frameworks (BrowserGym, WorkArena) and development platforms (AgentLab). My goal is to bridge foundational research and scalable tools to advance the field.

Previously, I completed my Ph.D. at the Quebec Artificial Intelligence Institute (Mila) under Professor Laurent Charlin. During my doctoral studies, I collaborated with DeepMind’s Continual Learning team led by Marc’Aurelio Ranzato, Amazon’s team under Alex Smola, and ElementAI prior to its integration with ServiceNow.

My Ph.D. research focused on building agents capable of accumulating and transferring knowledge across tasks, drawing from continual learning, transfer learning, and meta-learning. My work explored applications in language, vision, and reinforcement learning, emphasizing improvements in data and compute efficiency.

About the AI Alliance The AI Alliance is an international community of researchers, developers and organizational leaders committed to support and enhance open innovation across the AI technology landscape to accelerate progress, improve safety, security and trust in AI, and maximize benefits to people and society everywhere. Members of the AI Alliance believe that open innovation is essential to develop and achieve safe and responsible AI that benefit society rather than benefit a select few big players.

Join the community Sign up for the AI Alliance newsletter (check the website footer) and join our new AI Alliance Discord.

LLM-based web agents have recently made significant progress, but much of it has occurred in closed-source systems, widening the gap with open-source alternatives. Progress has been held back by two key challenges: first, a narrow focus on single-step tasks that overlooks the complexity of multi-step web interactions; and second, the high compute costs required to post-train LLM-based web agents.

To address this, we present the first statistically grounded study on compute allocation for LLM web-agent post-training. Our approach uses a two-stage pipeline, training a Llama 3.1 8B or QWEN 2.5 7B student to imitate a Llama 3.3 70B teacher or QWEN 2.5 72B via supervised fine-tuning (SFT), followed by on-policy reinforcement learning (GRPO).

We find this process highly sensitive to hyperparameter choices, making exhaustive sweeps impractical. To spare others from expensive trial-and-error, we sample 1,370 configurations and use bootstrapping to estimate effective hyperparameters. Our results show that combining SFT with on-policy RL consistently outperforms either approach alone on both WorkArena and MiniWob++. Further, this strategy requires only 55% of the compute to match the peak performance of pure SFT on MiniWob++, effectively pushing the compute-performance Pareto frontier, and is the only strategy that can close the gap with closed-source models.

Read the paper on ArXiv: How to Train Your LLM Web Agent: A Statistical Diagnosis (PDF)

About the speaker

I’m Massimo Caccia, Senior Research Scientist at ServiceNow Research, specializing in post-training methods for computer-use agents. I see computer use as the ultimate playground for testing agents, thanks to its ubiquity and diversity. My research involves conducting large-scale empirical studies to systematically evaluate trade-offs among different approaches and to develop practical know-how, with reinforcement learning being a particular focus.

As a core contributor to the web-agent research library ecosystem, I actively shape evaluation frameworks (BrowserGym, WorkArena) and development platforms (AgentLab). My goal is to bridge foundational research and scalable tools to advance the field.

Previously, I completed my Ph.D. at the Quebec Artificial Intelligence Institute (Mila) under Professor Laurent Charlin. During my doctoral studies, I collaborated with DeepMind’s Continual Learning team led by Marc’Aurelio Ranzato, Amazon’s team under Alex Smola, and ElementAI prior to its integration with ServiceNow.

My Ph.D. research focused on building agents capable of accumulating and transferring knowledge across tasks, drawing from continual learning, transfer learning, and meta-learning. My work explored applications in language, vision, and reinforcement learning, emphasizing improvements in data and compute efficiency.

About the AI Alliance The AI Alliance is an international community of researchers, developers and organizational leaders committed to support and enhance open innovation across the AI technology landscape to accelerate progress, improve safety, security and trust in AI, and maximize benefits to people and society everywhere. Members of the AI Alliance believe that open innovation is essential to develop and achieve safe and responsible AI that benefit society rather than benefit a select few big players.

Join the community Sign up for the AI Alliance newsletter (check the website footer) and join our new AI Alliance Discord.

LLM-based web agents have recently made significant progress, but much of it has occurred in closed-source systems, widening the gap with open-source alternatives. Progress has been held back by two key challenges: first, a narrow focus on single-step tasks that overlooks the complexity of multi-step web interactions; and second, the high compute costs required to post-train LLM-based web agents.

To address this, we present the first statistically grounded study on compute allocation for LLM web-agent post-training. Our approach uses a two-stage pipeline, training a Llama 3.1 8B or QWEN 2.5 7B student to imitate a Llama 3.3 70B teacher or QWEN 2.5 72B via supervised fine-tuning (SFT), followed by on-policy reinforcement learning (GRPO).

We find this process highly sensitive to hyperparameter choices, making exhaustive sweeps impractical. To spare others from expensive trial-and-error, we sample 1,370 configurations and use bootstrapping to estimate effective hyperparameters. Our results show that combining SFT with on-policy RL consistently outperforms either approach alone on both WorkArena and MiniWob++. Further, this strategy requires only 55% of the compute to match the peak performance of pure SFT on MiniWob++, effectively pushing the compute-performance Pareto frontier, and is the only strategy that can close the gap with closed-source models.

Read the paper on ArXiv: How to Train Your LLM Web Agent: A Statistical Diagnosis (PDF)

About the speaker

I’m Massimo Caccia, Senior Research Scientist at ServiceNow Research, specializing in post-training methods for computer-use agents. I see computer use as the ultimate playground for testing agents, thanks to its ubiquity and diversity. My research involves conducting large-scale empirical studies to systematically evaluate trade-offs among different approaches and to develop practical know-how, with reinforcement learning being a particular focus.

As a core contributor to the web-agent research library ecosystem, I actively shape evaluation frameworks (BrowserGym, WorkArena) and development platforms (AgentLab). My goal is to bridge foundational research and scalable tools to advance the field.

Previously, I completed my Ph.D. at the Quebec Artificial Intelligence Institute (Mila) under Professor Laurent Charlin. During my doctoral studies, I collaborated with DeepMind’s Continual Learning team led by Marc’Aurelio Ranzato, Amazon’s team under Alex Smola, and ElementAI prior to its integration with ServiceNow.

My Ph.D. research focused on building agents capable of accumulating and transferring knowledge across tasks, drawing from continual learning, transfer learning, and meta-learning. My work explored applications in language, vision, and reinforcement learning, emphasizing improvements in data and compute efficiency.

About the AI Alliance The AI Alliance is an international community of researchers, developers and organizational leaders committed to support and enhance open innovation across the AI technology landscape to accelerate progress, improve safety, security and trust in AI, and maximize benefits to people and society everywhere. Members of the AI Alliance believe that open innovation is essential to develop and achieve safe and responsible AI that benefit society rather than benefit a select few big players.

Join the community Sign up for the AI Alliance newsletter (check the website footer) and join our new AI Alliance Discord.

LLM-based web agents have recently made significant progress, but much of it has occurred in closed-source systems, widening the gap with open-source alternatives. Progress has been held back by two key challenges: first, a narrow focus on single-step tasks that overlooks the complexity of multi-step web interactions; and second, the high compute costs required to post-train LLM-based web agents.

To address this, we present the first statistically grounded study on compute allocation for LLM web-agent post-training. Our approach uses a two-stage pipeline, training a Llama 3.1 8B or QWEN 2.5 7B student to imitate a Llama 3.3 70B teacher or QWEN 2.5 72B via supervised fine-tuning (SFT), followed by on-policy reinforcement learning (GRPO).

We find this process highly sensitive to hyperparameter choices, making exhaustive sweeps impractical. To spare others from expensive trial-and-error, we sample 1,370 configurations and use bootstrapping to estimate effective hyperparameters. Our results show that combining SFT with on-policy RL consistently outperforms either approach alone on both WorkArena and MiniWob++. Further, this strategy requires only 55% of the compute to match the peak performance of pure SFT on MiniWob++, effectively pushing the compute-performance Pareto frontier, and is the only strategy that can close the gap with closed-source models.

Read the paper on ArXiv: How to Train Your LLM Web Agent: A Statistical Diagnosis (PDF)

About the speaker

I’m Massimo Caccia, Senior Research Scientist at ServiceNow Research, specializing in post-training methods for computer-use agents. I see computer use as the ultimate playground for testing agents, thanks to its ubiquity and diversity. My research involves conducting large-scale empirical studies to systematically evaluate trade-offs among different approaches and to develop practical know-how, with reinforcement learning being a particular focus.

As a core contributor to the web-agent research library ecosystem, I actively shape evaluation frameworks (BrowserGym, WorkArena) and development platforms (AgentLab). My goal is to bridge foundational research and scalable tools to advance the field.

Previously, I completed my Ph.D. at the Quebec Artificial Intelligence Institute (Mila) under Professor Laurent Charlin. During my doctoral studies, I collaborated with DeepMind’s Continual Learning team led by Marc’Aurelio Ranzato, Amazon’s team under Alex Smola, and ElementAI prior to its integration with ServiceNow.

My Ph.D. research focused on building agents capable of accumulating and transferring knowledge across tasks, drawing from continual learning, transfer learning, and meta-learning. My work explored applications in language, vision, and reinforcement learning, emphasizing improvements in data and compute efficiency.

About the AI Alliance The AI Alliance is an international community of researchers, developers and organizational leaders committed to support and enhance open innovation across the AI technology landscape to accelerate progress, improve safety, security and trust in AI, and maximize benefits to people and society everywhere. Members of the AI Alliance believe that open innovation is essential to develop and achieve safe and responsible AI that benefit society rather than benefit a select few big players.

Join the community Sign up for the AI Alliance newsletter (check the website footer) and join our new AI Alliance Discord.

LLM-based web agents have recently made significant progress, but much of it has occurred in closed-source systems, widening the gap with open-source alternatives. Progress has been held back by two key challenges: first, a narrow focus on single-step tasks that overlooks the complexity of multi-step web interactions; and second, the high compute costs required to post-train LLM-based web agents.

To address this, we present the first statistically grounded study on compute allocation for LLM web-agent post-training. Our approach uses a two-stage pipeline, training a Llama 3.1 8B or QWEN 2.5 7B student to imitate a Llama 3.3 70B teacher or QWEN 2.5 72B via supervised fine-tuning (SFT), followed by on-policy reinforcement learning (GRPO).

We find this process highly sensitive to hyperparameter choices, making exhaustive sweeps impractical. To spare others from expensive trial-and-error, we sample 1,370 configurations and use bootstrapping to estimate effective hyperparameters. Our results show that combining SFT with on-policy RL consistently outperforms either approach alone on both WorkArena and MiniWob++. Further, this strategy requires only 55% of the compute to match the peak performance of pure SFT on MiniWob++, effectively pushing the compute-performance Pareto frontier, and is the only strategy that can close the gap with closed-source models.

Read the paper on ArXiv: How to Train Your LLM Web Agent: A Statistical Diagnosis (PDF)

About the speaker

I’m Massimo Caccia, Senior Research Scientist at ServiceNow Research, specializing in post-training methods for computer-use agents. I see computer use as the ultimate playground for testing agents, thanks to its ubiquity and diversity. My research involves conducting large-scale empirical studies to systematically evaluate trade-offs among different approaches and to develop practical know-how, with reinforcement learning being a particular focus.

As a core contributor to the web-agent research library ecosystem, I actively shape evaluation frameworks (BrowserGym, WorkArena) and development platforms (AgentLab). My goal is to bridge foundational research and scalable tools to advance the field.

Previously, I completed my Ph.D. at the Quebec Artificial Intelligence Institute (Mila) under Professor Laurent Charlin. During my doctoral studies, I collaborated with DeepMind’s Continual Learning team led by Marc’Aurelio Ranzato, Amazon’s team under Alex Smola, and ElementAI prior to its integration with ServiceNow.

My Ph.D. research focused on building agents capable of accumulating and transferring knowledge across tasks, drawing from continual learning, transfer learning, and meta-learning. My work explored applications in language, vision, and reinforcement learning, emphasizing improvements in data and compute efficiency.

About the AI Alliance The AI Alliance is an international community of researchers, developers and organizational leaders committed to support and enhance open innovation across the AI technology landscape to accelerate progress, improve safety, security and trust in AI, and maximize benefits to people and society everywhere. Members of the AI Alliance believe that open innovation is essential to develop and achieve safe and responsible AI that benefit society rather than benefit a select few big players.

Join the community Sign up for the AI Alliance newsletter (check the website footer) and join our new AI Alliance Discord.

LLM-based web agents have recently made significant progress, but much of it has occurred in closed-source systems, widening the gap with open-source alternatives. Progress has been held back by two key challenges: first, a narrow focus on single-step tasks that overlooks the complexity of multi-step web interactions; and second, the high compute costs required to post-train LLM-based web agents.

To address this, we present the first statistically grounded study on compute allocation for LLM web-agent post-training. Our approach uses a two-stage pipeline, training a Llama 3.1 8B or QWEN 2.5 7B student to imitate a Llama 3.3 70B teacher or QWEN 2.5 72B via supervised fine-tuning (SFT), followed by on-policy reinforcement learning (GRPO).

We find this process highly sensitive to hyperparameter choices, making exhaustive sweeps impractical. To spare others from expensive trial-and-error, we sample 1,370 configurations and use bootstrapping to estimate effective hyperparameters. Our results show that combining SFT with on-policy RL consistently outperforms either approach alone on both WorkArena and MiniWob++. Further, this strategy requires only 55% of the compute to match the peak performance of pure SFT on MiniWob++, effectively pushing the compute-performance Pareto frontier, and is the only strategy that can close the gap with closed-source models.

Read the paper on ArXiv: How to Train Your LLM Web Agent: A Statistical Diagnosis (PDF)

About the speaker

I’m Massimo Caccia, Senior Research Scientist at ServiceNow Research, specializing in post-training methods for computer-use agents. I see computer use as the ultimate playground for testing agents, thanks to its ubiquity and diversity. My research involves conducting large-scale empirical studies to systematically evaluate trade-offs among different approaches and to develop practical know-how, with reinforcement learning being a particular focus.

As a core contributor to the web-agent research library ecosystem, I actively shape evaluation frameworks (BrowserGym, WorkArena) and development platforms (AgentLab). My goal is to bridge foundational research and scalable tools to advance the field.

Previously, I completed my Ph.D. at the Quebec Artificial Intelligence Institute (Mila) under Professor Laurent Charlin. During my doctoral studies, I collaborated with DeepMind’s Continual Learning team led by Marc’Aurelio Ranzato, Amazon’s team under Alex Smola, and ElementAI prior to its integration with ServiceNow.

My Ph.D. research focused on building agents capable of accumulating and transferring knowledge across tasks, drawing from continual learning, transfer learning, and meta-learning. My work explored applications in language, vision, and reinforcement learning, emphasizing improvements in data and compute efficiency.

About the AI Alliance The AI Alliance is an international community of researchers, developers and organizational leaders committed to support and enhance open innovation across the AI technology landscape to accelerate progress, improve safety, security and trust in AI, and maximize benefits to people and society everywhere. Members of the AI Alliance believe that open innovation is essential to develop and achieve safe and responsible AI that benefit society rather than benefit a select few big players.

Join the community Sign up for the AI Alliance newsletter (check the website footer) and join our new AI Alliance Discord.

LLM-based web agents have recently made significant progress, but much of it has occurred in closed-source systems, widening the gap with open-source alternatives. Progress has been held back by two key challenges: first, a narrow focus on single-step tasks that overlooks the complexity of multi-step web interactions; and second, the high compute costs required to post-train LLM-based web agents.

To address this, we present the first statistically grounded study on compute allocation for LLM web-agent post-training. Our approach uses a two-stage pipeline, training a Llama 3.1 8B or QWEN 2.5 7B student to imitate a Llama 3.3 70B teacher or QWEN 2.5 72B via supervised fine-tuning (SFT), followed by on-policy reinforcement learning (GRPO).

We find this process highly sensitive to hyperparameter choices, making exhaustive sweeps impractical. To spare others from expensive trial-and-error, we sample 1,370 configurations and use bootstrapping to estimate effective hyperparameters. Our results show that combining SFT with on-policy RL consistently outperforms either approach alone on both WorkArena and MiniWob++. Further, this strategy requires only 55% of the compute to match the peak performance of pure SFT on MiniWob++, effectively pushing the compute-performance Pareto frontier, and is the only strategy that can close the gap with closed-source models.

Read the paper on ArXiv: How to Train Your LLM Web Agent: A Statistical Diagnosis (PDF)

About the speaker

I’m Massimo Caccia, Senior Research Scientist at ServiceNow Research, specializing in post-training methods for computer-use agents. I see computer use as the ultimate playground for testing agents, thanks to its ubiquity and diversity. My research involves conducting large-scale empirical studies to systematically evaluate trade-offs among different approaches and to develop practical know-how, with reinforcement learning being a particular focus.

As a core contributor to the web-agent research library ecosystem, I actively shape evaluation frameworks (BrowserGym, WorkArena) and development platforms (AgentLab). My goal is to bridge foundational research and scalable tools to advance the field.

Previously, I completed my Ph.D. at the Quebec Artificial Intelligence Institute (Mila) under Professor Laurent Charlin. During my doctoral studies, I collaborated with DeepMind’s Continual Learning team led by Marc’Aurelio Ranzato, Amazon’s team under Alex Smola, and ElementAI prior to its integration with ServiceNow.

My Ph.D. research focused on building agents capable of accumulating and transferring knowledge across tasks, drawing from continual learning, transfer learning, and meta-learning. My work explored applications in language, vision, and reinforcement learning, emphasizing improvements in data and compute efficiency.

About the AI Alliance The AI Alliance is an international community of researchers, developers and organizational leaders committed to support and enhance open innovation across the AI technology landscape to accelerate progress, improve safety, security and trust in AI, and maximize benefits to people and society everywhere. Members of the AI Alliance believe that open innovation is essential to develop and achieve safe and responsible AI that benefit society rather than benefit a select few big players.

Join the community Sign up for the AI Alliance newsletter (check the website footer) and join our new AI Alliance Discord.

Summary In this episode of the Data Engineering Podcast Alex Albu, tech lead for AI initiatives at Starburst, talks about integrating AI workloads with the lakehouse architecture. From his software engineering roots to leading data engineering efforts, Alex shares insights on enhancing Starburst's platform to support AI applications, including an AI agent for data exploration and using AI for metadata enrichment and workload optimization. He discusses the challenges of integrating AI with data systems, innovations like SQL functions for AI tasks and vector databases, and the limitations of traditional architectures in handling AI workloads. Alex also shares his vision for the future of Starburst, including support for new data formats and AI-driven data exploration tools.

Announcements Hello and welcome to the Data Engineering Podcast, the show about modern data managementData migrations are brutal. They drag on for months—sometimes years—burning through resources and crushing team morale. Datafold's AI-powered Migration Agent changes all that. Their unique combination of AI code translation and automated data validation has helped companies complete migrations up to 10 times faster than manual approaches. And they're so confident in their solution, they'll actually guarantee your timeline in writing. Ready to turn your year-long migration into weeks? Visit dataengineeringpodcast.com/datafold today for the details.This is a pharmaceutical Ad for Soda Data Quality. Do you suffer from chronic dashboard distrust? Are broken pipelines and silent schema changes wreaking havoc on your analytics? You may be experiencing symptoms of Undiagnosed Data Quality Syndrome — also known as UDQS. Ask your data team about Soda. With Soda Metrics Observability, you can track the health of your KPIs and metrics across the business — automatically detecting anomalies before your CEO does. It’s 70% more accurate than industry benchmarks, and the fastest in the category, analyzing 1.1 billion rows in just 64 seconds. And with Collaborative Data Contracts, engineers and business can finally agree on what “done” looks like — so you can stop fighting over column names, and start trusting your data again.Whether you’re a data engineer, analytics lead, or just someone who cries when a dashboard flatlines, Soda may be right for you. Side effects of implementing Soda may include: Increased trust in your metrics, reduced late-night Slack emergencies, spontaneous high-fives across departments, fewer meetings and less back-and-forth with business stakeholders, and in rare cases, a newfound love of data. Sign up today to get a chance to win a $1000+ custom mechanical keyboard. Visit dataengineeringpodcast.com/soda to sign up and follow Soda’s launch week. It starts June 9th. This episode is brought to you by Coresignal, your go-to source for high-quality public web data to power best-in-class AI products. Instead of spending time collecting, cleaning, and enriching data in-house, use ready-made multi-source B2B data that can be smoothly integrated into your systems via APIs or as datasets. With over 3 billion data records from 15+ online sources, Coresignal delivers high-quality data on companies, employees, and jobs. It is powering decision-making for more than 700 companies across AI, investment, HR tech, sales tech, and market intelligence industries. A founding member of the Ethical Web Data Collection Initiative, Coresignal stands out not only for its data quality but also for its commitment to responsible data collection practices. Recognized as the top data provider by Datarade for two consecutive years, Coresignal is the go-to partner for those who need fresh, accurate, and ethically sourced B2B data at scale. Discover how Coresignal's data can enhance your AI platforms. Visit dataengineeringpodcast.com/coresignal to start your free 14-day trial.Your host is Tobias Macey and today I'm interviewing Alex Albu about how Starburst is extending the lakehouse to support AI workloadsInterview IntroductionHow did you get involved in the area of data management?Can you start by outlining the interaction points of AI with the types of data workflows that you are supporting with Starburst?What are some of the limitations of warehouse and lakehouse systems when it comes to supporting AI systems?What are the points of friction for engineers who are trying to employ LLMs in the work of maintaining a lakehouse environment?Methods such as tool use (exemplified by MCP) are a means of bolting on AI models to systems like Trino. What are some of the ways that is insufficient or cumbersome?Can you describe the technical implementation of the AI-oriented features that you have incorporated into the Starburst platform?What are the foundational architectural modifications that you had to make to enable those capabilities?For the vector storage and indexing, what modifications did you have to make to iceberg?What was your reasoning for not using a format like Lance?For teams who are using Starburst and your new AI features, what are some examples of the workflows that they can expect?What new capabilities are enabled by virtue of embedding AI features into the interface to the lakehouse?What are the most interesting, innovative, or unexpected ways that you have seen Starburst AI features used?What are the most interesting, unexpected, or challenging lessons that you have learned while working on AI features for Starburst?When is Starburst/lakehouse the wrong choice for a given AI use case?What do you have planned for the future of AI on Starburst?Contact Info LinkedInParting Question From your perspective, what is the biggest gap in the tooling or technology for data management today?Closing Announcements Thank you for listening! Don't forget to check out our other shows. Podcast.init covers the Python language, its community, and the innovative ways it is being used. The AI Engineering Podcast is your guide to the fast-moving world of building AI systems.Visit the site to subscribe to the show, sign up for the mailing list, and read the show notes.If you've learned something or tried out a project from the show then tell us about it! Email [email protected] with your story.Links StarburstPodcast EpisodeAWS AthenaMCP == Model Context ProtocolLLM Tool UseVector EmbeddingsRAG == Retrieval Augmented GenerationAI Engineering Podcast EpisodeStarburst Data ProductsLanceLanceDBParquetORCpgvectorStarburst IcehouseThe intro and outro music is from The Hug by The Freak Fandango Orchestra / CC BY-SA

What are dbt meetups? dbt Meetups are networking events open to all folks working with data! Talks predominantly focus on community members' experience with dbt, however, you'll catch presentations on broader topics such as analytics engineering, data stacks, data ops, modeling, testing, and team structures.

🤝Organizer: Sarah Hennig (Taod Consulting) 🏠Venue Host: Taod Consulting, GmbH, Oskar-Jäger-Str. 173, 50825 Köln 🍕Catering: light bites and refreshments

To attend, please read the Health and Safety Policy and Terms of Participation: https://www.getdbt.com/legal/health-and-safety-policy

Agenda

• 6:00 PM \| Check-in / Registration
• 6:30 PM \| Welcome Remarks & Housekeeping by taod
• 7:00 PM \| Stephan Durry (dbt Labs) - "What's new at dbt Labs"
• 7:15 PM \| First Talk: Sönke Maibach (Data Engineer\, taod)\, - "Lessons on migrating legacy stack to dbt reliably and cheaply"
• 7:45 PM \| Second Talk: Alex Rupp (VP Data & BI\, Schüttflix) - "Pairing dbt and Slack Bots - for data-driven communication"
• 8:15 PM \| Networking & Reception

➡️ Join the dbt Slack community: https://www.getdbt.com/community/ 🤝For the best Meetup experience, make sure to join the #local-northern-germany channel in dbt Slack (https://slack.getdbt.com/). ---------------------------------- dbt is the standard in data transformation, used by over 40,000 organizations worldwide. Through the application of software engineering best practices like modularity, version control, testing, and documentation, dbt’s analytics engineering workflow helps teams work more efficiently to produce data the entire organization can trust. Learn more: https://www.getdbt.com

What are dbt meetups? dbt Meetups are networking events open to all folks working with data! Talks predominantly focus on community members' experience with dbt, however, you'll catch presentations on broader topics such as analytics engineering, data stacks, data ops, modeling, testing, and team structures.

🤝Organizer: dbt Labs (Amada Echeverría, Community Team) 🏠Venue/Host: Snowflake Office (Room: the Lodge on the 5th floor) @ One Crown Place, London EC2A 4EF, UK 🍕Catering: Refreshments will be provided

To attend, please read the Health and Safety Policy and Terms of Participation: https://www.getdbt.com/legal/health-and-safety-policy

📝Agenda

• 5:00 - 5:40 \| Check in/Registration (40 min)
• 5:40 - 5:45 \| Welcome Remarks by our meetup host\, Snowflake (5 min) and dbt Labs’ Community Team
• 5:45 - 6:05 \| Presentation 1: Kshitij Aranke (Senior Software Engineer @ dbt Labs) on ‘Hands-on with our latest innovations in dbt Core 1.10” (20 min)
• 6:05 - 6:15 \| Presentation 1 Q&A (10 min)
• 6:15 - 6:35 \| Fireside chat with Tristan Handy (Founder & CEO @ dbt Labs) and Alex Lambert (Director Solution Engineering UK @ Snowflake) on The Analytics Development Lifecycle (20 min)
• 6:35 - 6:45 \| Q&A for Tristan and Alex (10 min)
• 6:45 -6:50 \| Closing Remarks by Snowflake and dbt Labs’ Community Team (5 min)
• 6:50 - 8:00 \| Reception (Please be mindful that we need to leave the space by 8:30)

EVENT DETAILS: The doors open at 5pm. Presentations begin at 5:40pm. Food and refreshments will be provided. The venue has capacity limits, so please update your RSVP status on the Meetup to "Not Going" if you are unable to attend.

DIRECTIONS: Please bring an ID. The Snowflake office is located on the 5th and 6th floor of a multi-tenant building. Upon arrival visitors are greeted at ground floor reception and asked to check-in and sign Snowflake’s NDA via the Envoy check-in system. Visitors are either escorted to the lift area or await the host to pick them up.

➡️ Join the dbt Slack community: https://www.getdbt.com/community/ 🤝For the best Meetup experience, make sure to join the #local-ldn channel in dbt Slack (https://slack.getdbt.com/). ---------------------------------- dbt is the standard in data transformation, used by over 40,000 organizations worldwide. Through the application of software engineering best practices like modularity, version control, testing, and documentation, dbt’s analytics engineering workflow helps teams work more efficiently to produce data the entire organization can trust. Learn more: https://www.getdbt.com/

Government Blockchain Association Public Webinar "Blockchain Technology’s Role in Economic Inclusion"

Presented by GBA Cryptocurrency Working Group (CWG) Working Group & eazybot.com Thursday Mar 27, 2025, 12:00 PM - 1:00 PM EDT

Register for the full series here: CWG 03-27 – from the PDF Source https://eazybot.co/CWG032725u

"Blockchain Technology’s Role in Economic Inclusion"

As blockchain continues to reshape global financial ecosystems, this session will explore its ability to foster financial inclusion, expand banking access, and create new economic opportunities for underserved populations. Your expertise will play a critical role in shedding light on the opportunities, challenges, and evolving policies shaping this space.

Moderated by Charles Anderson of eazybot.com with guest speakers" Ryan Munn, Founder, Interchain, LLC Alex Short, Project Director (PMSCP), United States Global Operations Leader (GBA) Dr. Josina Rodriguez, Academic Advisory Body Member for the International Association for Trusted Blockchain Applications Dr. Christopher Smithmyer, Vice President of International Affairs, Brāv Online Conflict Management and Co-Founder of Black Wallet Limited

To ensure a robust and insightful discussion, we encourage you to consider the following topics: ✅ The Role of Blockchain in Expanding Financial Access How blockchain-based solutions are removing barriers for the unbanked and underbanked. The impact of decentralized finance (DeFi) and blockchain-based remittances. Case studies of successful blockchain-driven financial inclusion initiatives. ✅ Government and Regulatory Developments in Blockchain and Economic Inclusion The implications of the Executive Order on Strengthening American Leadership in Digital Financial Technology (January 23, 2025) on economic inclusion. 🔗 Read the Executive Order 🔗 White House Fact Sheet ✅ The Strategic Bitcoin Reserve: Global and Domestic Impact The growing state and federal discussions around Bitcoin as a strategic reserve asset. The potential benefits and risks of nations adopting Bitcoin as a financial reserve. 🔗 Forbes: Nations & States Turn to Bitcoin as a Strategic Reserve 🔗 Bitcoin Reserve Legislation Expands to 15 States ✅ Federal Oversight and the Crypto Czar’s First Public Address Key takeaways from the U.S. Crypto Czar’s speech regarding financial regulation and economic inclusion. How regulatory changes could impact blockchain’s role in domestic and international finance. 🔗 CCN: Crypto Czar's Bitcoin Reserve Strategy ✅ Banking and Crypto: Evolving Regulatory Perspectives The Office of the Comptroller of the Currency (OCC) clarification on banks engaging with cryptocurrency. 🔗 OCC Clarifies Bank Authority to Engage in Certain Cryptocurrency Activities The SEC’s confirmation that Bitcoin mining is not considered a security under U.S. law and its implications for future regulation. 🔗 SEC Confirms Proof-of-Work Mining is Not Subject to Securities Laws ✅ Current Events and Market Trends in Crypto Recent global regulatory updates, market movements, and economic policies impacting blockchain-based financial solutions. 🔗 Cointelegraph Latest News 🔗 CryptoNews Updates

Join Webinar at https://eazybot.co/CWG032725u Contact Bob Miko, [email protected] 203 378 2803

About GBA

The Government Blockchain Association (GBA) is a nonprofit organization committed to advancing blockchain technology standards, facilitating industry education, and ensuring a trusted, secure ecosystem for blockchain solutions. For more information, visit https://gbaglobal.org/.

For more information contact Bob Miko, [email protected] 203 378 2803

--

Bob Miko GBA Director of Public Relations Editor in Chief/Producer Pacific Dialogue 203 378 2803 [email protected]

We're pleased to present Alex Weisberger on Performal: Formal Verification of Latency Properties for Distributed Systems (Read the Paper)

Performal is an approach for formally reasoning about latency in distributed systems. It provides a way to state and prove worst-case latency bounds in a way that abstracts over the complexity of real-world infrastructure. To deal with the messy details of physical reality, it provides a way to pair this with performance measurements so that end-to-end latency can be accurately estimated.

Alex Weisberger is a full-stack engineer currently working at Datadog on the Database Monitoring product. At work, he's been working on database health analysis and recommendations. His extracurricular tech interests are around applying formal methods to day-to-day work: things like property-based testing, model-based testing, TLA+, and simulation. His weekends are mostly filled with Nerf fights, Super Smash Bros., and various sports with his kids.

---

⚠️ Required: You must have your real name on your account and provide a photo ID at the entrance to attend, per the venue rules. If you are not on the list, you will not be admitted.

🚔 Reminder: Papers We Love has a code of conduct. Breaching the CoC is grounds to be ejected from the meetup at the organizers' discretion.

📹 The event will be recorded and made available 1-2 weeks afterwards.

💬 Join us on the Papers We Love Discord - https://discord.gg/6gupsBg4qp

Venue:

Datadog 620 8th Ave, 45th Floor New York, NY 10018 USA Doors open at 6:30pm EST Note: Enter the building on 8th Avenue and go to the reception desk on your left

⚠️ Required: You must have your real name on your Meetup account and provide a photo ID at the entrance to attend, per the venue rules. If you are not on the list, you will not be admitted.

We're pleased to present Alex Kesling on Apache Arrow DataFusion: a Fast, Embeddable, Modular Analytic Query Engine (Read the Paper)

DataFusion represents a new category of database technologies: "deconstructed databases". It's a relational engine on which you can graft any language frontend (with SQL provided out of the box) and point it at any data you want.

How could one technology possibly do the same work as SQLite, PostgreSQL, Pandas, DuckDB, what-have-you all in one package? Let's read the paper and talk about it.

Alex Kesling is the cofounder of Empathic, a new company exploring the intersection of enterprise data infrastructure and agentic AI. He previously worked on Google's Search and Knowledge Graph infrastructure, VR/AR tools for teachers, and a plethora of consulting projects. Alex loves systems design, time series visualization, distributed systems, and that one turkey sandwich the day after Thanksgiving.

---

⚠️ Required: You must have your real name on your account and provide a photo ID at the entrance to attend, per the venue rules. If you are not on the list, you will not be admitted.

🚔 Reminder: Papers We Love has a code of conduct. Breaching the CoC is grounds to be ejected from the meetup at the organizers' discretion.

📹 The event will be recorded and made available 1-2 weeks afterwards.

💬 Join us on the Papers We Love Discord - https://discord.gg/6gupsBg4qp

Venue:

Datadog 620 8th Ave, 45th Floor New York, NY 10018 USA Doors open at 6:30pm EST Note: Enter the building on 8th Avenue and go to the reception desk on your left

⚠️ Required: You must have your real name on your Meetup account and provide a photo ID at the entrance to attend, per the venue rules. If you are not on the list, you will not be admitted.

Discover how to cut complexity of your dbt data pipelines with serverless DuckDB while improving performance and drastically reducing costs. This session covers practical strategies for cutting complexity and expenses in data flows while enjoying a more ergonomic and frictionless workflow. Learn how adopting a DuckDB-based architecture can streamline your operations, enhance developer experience, and boost efficiency.

Speaker: Alex Monahan Forward Deployed Software Engineer MotherDuck

Read the blog to learn about the latest dbt Cloud features announced at Coalesce, designed to help organizations embrace analytics best practices at scale https://www.getdbt.com/blog/coalesce-2024-product-announcements

Welcome to the PyData Berlin September meetup!

We would like to welcome you all starting from 18:30. There will be food and drinks. The talks begin around 19.00.

Please provide your first and last name for the registration because this is required for the venue's entry policy. If you cannot attend, please cancel your spot so others can join as the space is limited.

Host: SPICED Academy is excited to welcome you for this month's version of PyData. At SPICED Academy, we connect the next tech success stories to their futures. Our intensive bootcamps in Data Science, Data Analytics, UX/UI, Java Development and Web Development teach the most up-to-date, in-demand skills on the market. **************************************************************************

The Lineup for the evening

Talk 1: FAir Interpolation Method (FAIM) for Trade-Offs between Mutually Exclusive Algorithmic Fairness Criteria in Machine Learning and Law? Abstract: Ensuring fairness in AI is crucial but challenging, as different fairness criteria often conflict. This talk introduces the FAir Interpolation Method (FAIM), a novel algorithm using optimal transport to balance competing fairness measures, such as ‘calibration within groups’ and ‘balance for the positive/negative class’. FAIM aims to adjust unfair predictions to meet a weighted combination of fairness standards, and its effectiveness is demonstrated with synthetic credit score data, COMPAS data, and e-commerce recommendations. FAIM could help align AI systems with legal standards, including recent EU regulations. Speaker: Alex Loosley is a Responsible AI Scientist and Engineer with experience in developing algorithms for crop protection, document intelligence, and fashion fit recommendations. Outside of his AI work, he referees professional ice hockey in Germany.

Talk 2: AI on a Microbudget - Methods of Machine Learning Miniaturization? Abstract: Current progress in AI has seen remarkable capabilities emerging from simple prediction tasks – if we scale them massively. Surprisingly, we get sparks of reasoning and intelligence in a model that was trained to do little more than masked word prediction. Since that realization the AI field has pursued ever larger models, trained at “eye-watering” cost. If scaling is all you need – does it follow that, in practice, money is all you need? In this talk we explore ideas for the rest of us, the GPU-poor. Taking examples from language processing and computer vision, we’ll show you how to make do with less – less computing power, less person power, less data – while still building powerful models. We will introduce a set of methods and open source tools for the efficient reuse and miniaturization of models, including transfer learning and fine-tuning, knowledge distillation, and model quantization. Our talk aims to provide an overview for ML practitioners, draws from our combined project experience, and is accompanied by a repository of code examples to get you started with building AI on a microbudget. Speaker: Katharina Rasch is a data scientist and computer vision engineer with a PhD in Computer Science from KTH Stockholm. She currently freelances in Berlin and also works as a teacher. Christian Staudt is a data scientist with 8 years of freelance experience. He focuses on machine learning from prototype to deployment, contributes to open source and has organized PyData community events.

Lightning talks There will be slots for 2-3 Lightning Talks (3-5 Minutes for each). Kindly let us know if you would like to present something at the start of the meetup :)

*** NumFOCUS Code of Conduct THE SHORT VERSION Be kind to others. Do not insult or put down others. Behave professionally. Remember that harassment and sexist, racist, or exclusionary jokes are not appropriate for NumFOCUS. All communication should be appropriate for a professional audience including people of many different backgrounds. Sexual language and imagery are not appropriate. NumFOCUS is dedicated to providing a harassment-free community for everyone, regardless of gender, sexual orientation, gender identity, and expression, disability, physical appearance, body size, race, or religion. We do not tolerate harassment of community members in any form. Thank you for helping make this a welcoming, friendly community for all. If you haven't yet, please read the detailed version here: https://numfocus.org/code-of-conduct

AGENDA 18.00 – 18:20 Meet & Greet

18:20 - 19:10: I read the Databricks website and I still don't know what you do - by Holly Smith

Databricks is a Data Intelligence Platform, and if you don't know what that means this talk is for you. We'll go through common data problems seen in every industry, take a buzzword free look at what Databricks does, what it's truly capable of and how it can benefit you in your day jobs.

BIO

Holly Smith is a multi award winning Data & AI expert who has over a decade of experience working with Data & AI teams in a variety of capacities from individual contributors all the way up to leadership. In her role as Staff Developer Advocate at Databricks she has worked with many multi national companies as they embark on their journey to the cutting edge of data. She is a renowned public speaker, teacher and minority in tech advocate and works with non profits through Datakind UK.

19:10 - 19:40 Pizza and Networking

19:40 - 20:30 Ensuring Code Quality with Unit Testing, Linting, and Automation with Databricks by Alex Cole

• In the fast-paced world of software development, maintaining high code quality is crucial. This session will delve into the essential practices of unit testing, linting, and automation, and how they can transform your development workflow. We’ll cover: • Unit Testing: Understanding the importance of unit tests, how to write effective tests, and best practices for maintaining them. • Linting: How linting tools can help you catch errors early, enforce coding standards, and improve code readability. • Automation: Integrating unit testing and linting into your CI/CD pipeline to ensure continuous quality and reduce manual effort. By the end of this session, you’ll have a solid grasp of how these practices can enhance your code quality, streamline your development process, and ultimately lead to more robust and maintainable software.

BIO

• Alex Cole is a Senior Architect specializing in data projects. A Databricks Solution Architect Champion and a leader in the Databricks Community of Practice at Cognizant, he holds multiple Azure certifications in solutions architecture, data engineering, data science, and AI engineering, along with the Advanced Databricks Data Engineering certification. Alex has made contributions to the Apache Commons project with his Java code and has actively participated in open-source projects, including Databricks Labs and the Open Source Unity Catalog Project. Beyond his professional work, he volunteers as a CIO for a local non-profit organisation. Alex blogs at www.alexcole.net

Hosted and Sponsored by BJSS - www.bjss.com - a leading technology and engineering consultancy for business.

Pizzas and Drinks during the event - thanks BJSS!

-------------- About BJSS - https://www.bjss.com/ BJSS is a leading technology and engineering consultancy for business. Trusted by our clients, we collaborate with some of the world’s leading organisations to deliver complex, innovative technology, engineering, and industry solutions that millions of people use every day. ---- Photos We ask that you do NOT take photos at this meetup. We will invite people to be included in a group photo/s during the event. Speakers will let you know if it's okay to photograph their presentation (excluding other attendees). You may see organisers taking photos during the talks. These will be of speakers, if they have agreed to this, and will not include faces of attendees.

In this session we will explore some of AutoGen foundational concepts such as enhanced inference, code executors, and more advanced concepts such as Agents and Agent flows. The session will leverage code both in Jupyter Notebooks and the Autogen Studio code.

Session objectives: * Basic understanding of some of the AutoGen's foundational and intermediate concepts * Basic understanding of the AutoGen Studio code * Building applications with AutoGen

Session Repo: https://github.com/microsoft/autogen/tree/main/samples/apps/autogen-studio

Presented by Alex Morales, Customer Engineering Manager at Microsoft

** Part of RAGHack, a free global hackathon to develop RAG applications. Join at https://aka.ms/raghack **

📌 Check out the RAGHack 2024 series here!

Pre-requisites: - Read the official rules and join the hack at https://aka.ms/raghack. No Purchase Necessary. Must be 18+ to enter. Contest ends 9/16/24.

• Want more hands-on RAG training? Visit the Reactor series home page to see all the RAGHack 2024 sessions!

In this session we will explore some of AutoGen foundational concepts such as enhanced inference, code executors, and more advanced concepts such as Agents and Agent flows. The session will leverage code both in Jupyter Notebooks and the Autogen Studio code.

Session objectives: * Basic understanding of some of the AutoGen's foundational and intermediate concepts * Basic understanding of the AutoGen Studio code * Building applications with AutoGen

Session Repo: https://github.com/microsoft/autogen/tree/main/samples/apps/autogen-studio

Presented by Alex Morales, Customer Engineering Manager at Microsoft

** Part of RAGHack, a free global hackathon to develop RAG applications. Join at https://aka.ms/raghack **

📌 Check out the RAGHack 2024 series here!

Pre-requisites: - Read the official rules and join the hack at https://aka.ms/raghack. No Purchase Necessary. Must be 18+ to enter. Contest ends 9/16/24.

• Want more hands-on RAG training? Visit the Reactor series home page to see all the RAGHack 2024 sessions!

Mark your calendars for Tuesday 9th July for our 16th OpenUK London community meetup we will be celebrating ten years of Kubernetes by talking to some of the UK's cloud native maintainers! Let's answer what's a cloud native mindset and how can you apply that to an open source community? What are some of the challenges?

SAVE THE DATE!

6:30 pm: Arrive, grab a bev, a slice and a chat. Scope out our swag table (bring pieces to contribute or trade!) 7:00 pm: Panel including Syntasso's Paula Kennedy, Otterize's Nicolas Vermandé, Isovalent at Cisco's Carla Gaggini and Akamai's Alex Chircop! We will be discussing how cloud native can enable delivering safe and high quality software and value to customers without slowing down – and how does open source software and community drive this? (This is the part that is livestreamed) 7:40 pm: TBA but it'll include a breakout that gets people talking to each other! 8:15-8:30 pm: Cleanup and see-ya-next-time. This will be our last meetup until September so not to be missed!

Please remember to bring your I.D. if you are attending in person, as it is required to access the venue. Note the venue does not permit e-scooters or bicycles within the building.

Thank you to our sponsors Avanade for providing this great space and yummy pizza! By attending this or any OpenUK event, you are adhering to our Code of Respect and OpenUK's Competition Policy. Please read ahead.

If you would like to hear more about OpenUK's other events, you can join our Newsletter. After attending, look out for an email with an invite to our burgeoning Slack community as a way to keep in touch and support each other between activities.

Registration Link: https://www.eventbrite.com/e/apples-vision-pro-vs-metas-quest-3-round-table-eugene-kent-alex-cix-tickets-846887803217?aff=meetup

This session directly follows the more technical/dev-focused event:

Port a Game from the Quest Store to the Apple visionOS App Store

Lecture Overview & Key Takeaways

This session is an Open Round Table discussing and comparing the Meta Quest 3 with the Apple Vision Pro Participating Thought Leaders:

• Eugene Chung (CEO, Penrose Studios)
• Amanda Watson (Former VR Software Engineer, Oculus)
• Erin Jerri Pañgilinan (Founder, Stealth)
• Daniel Sproll (CEO, Realities IO)
• Shahriar Shahrabi (Creative Director, Realities IO)
• Alex Coulombe (CEO, Agile Lens)
• Cix Liv (Founder, Vi)
• Fabian Quosdorf (Founder, mixed.world)

This is an interactive session. Feel free to ask any questions to the speakers.

Speaker

Eugene Chung is a globally recognized Blockchain Metaverse influencer, filmmaker, and investor as well as Founder & Chief Executive Officer of Penrose Studios. His latest work, “Arden's Wake", won the first Lion awarded for Best VR at the 74th Venice International Film Festival. Previously, he was Head of Film and Media and Co-Founder of Story Studio at Oculus VR, which was later acquired by Facebook for $2 billion. Prior to Oculus, he was in production at Pixar Animation Studios, a private equity investor at Warburg Pincus and an investment banker at Morgan Stanley. Connect with Eugene here: https://www.linkedin.com/in/eugeneykchung/

Amanda Watson is an XR developer and low-level systems hacker, best known for her work on Quest Air Link. Starting her career at post-acquisition Oculus, Amanda worked closely with the original team, gaining a deep understanding of the XR runtime and pipeline. This expertise led to her being on the foundational teams of many popular and well-loved VR products, including Oculus Quest, Quest Link, and Air Link. In October 2022, Amanda embarked on a new journey as an independent XR developer, motivated by the potential of new HMDs and the expanding XR landscape. Connect with Amanda here: https://www.linkedin.com/in/amwatson/

Erin Jerri Pañgilinan is a leading voice in emerging technology as an internationally acclaimed author, software engineer and computational designer, and educator in spatial computing and AI. Her book O’Reilly Media book, Creating Augmented and Virtual Realities: Theory and Practice for Next-Generation of Spatial Computing topped the must-read list for VR books in 2019, and debuted #1 in Game Programming on Amazon. The book has been translated into Chinese, Korean, and distributed in over 2 dozen countries. Currently, she is working on her next two books and releasing cross-platform productivity applications. Connect with Erin here: https://www.linkedin.com/in/erinjerri/

Alex Coulombe is an XR architect, and the founder of Agile Lens, a creative studio known for immersive design. His immersive storytelling work is acclaimed, especially within live theater. He introduced Heavenue, a high-fidelity cloud platform, that connects performers and metaverse audiences. He has taught VR performance at NYU Tisch, Unity3D at Syracuse University, and now instructs on Unreal Engine, as an authorized partner of Epic Games. Connect with Alex here: https://www.linkedin.com/in/alexcoulombe/

Cix Liv has been at the forefront of merging fitness and gaming with augmented and virtual reality technologies. Initiating his journey with LIV in 2016, he revolutionized VR live streaming. His venture into YUR in 2019 introduced fitness gaming for enthusiasts, followed by REK in 2022, which brought mixed reality sports to the forefront. His most recent endeavor, Vi, continues to explore further the potential of fitness gaming since May 2023. Connect with Cix here: https://www.linkedin.com/in/cix/