LLMday is a worldwide series of community events for engineers building with large language models, AI agents and machine learning. Across cities around the world, we bring together practitioners working on AI-powered products and systems to share real-world experience, learn from each other and explore how software engineering is changing in the age of AI.
Dr. Luis Bonavita 1266
11300 Montevideo, Departamento de Montevideo, Uruguay
Sponsors & Partners
Want to become a sponsor? Get in touch!
Thalita Regina Terpilauskas de Sa
Riverstone Insights - Iowa
AI Isn’t Magic—But It Can Deliver Remarkable Results
Abstract
Artificial intelligence is often presented as a magical solution capable of solving any business problem. However, many organizations begin their AI initiatives without a clearly defined use case, reliable data, or a realistic understanding of what the technology can deliver. This talk explores two of the most common challenges in enterprise AI adoption: the gap between business expectations and technical reality, and the lack of structured, consistent, and contextualized data. It also discusses the role of technical professionals in transforming broad AI ambitions into measurable and achievable solutions.
Finally, I will present a real-world predictive analytics project developed for the Brazilian consortium market. The case demonstrates how understanding the business problem, preparing the data, and selecting the right technical approach increased prediction accuracy from approximately 55% to 91.5%, turning an abstract AI initiative into a practical decision-support solution.
Bio
Thalita Regina Terpilauskas de Sa is a Business Intelligence, Data Analytics, and Automation professional with over 12 years of experience developing data-driven solutions for international organizations. Throughout her career, she has worked at leading global companies, including Louis Dreyfus Company, COFCO International, and Cargill. She is the founder and CEO of Riverstone Insights, a US-based data and AI consulting company that provides Business Intelligence, data analytics, automation, and AI solutions to clients such as HP, Italkraft, and DirectCon. Her expertise includes Power BI, SQL, Python, machine learning, predictive analytics, and business process automation. She recently completed the MIT IDSS Data Science and Machine Learning Program and focuses on transforming complex business challenges into practical, measurable technology solutions.
Marcelo B San Martin
Blizzard Entertainment
The Great AI Role Rewrite
Abstract
AI is no longer only automating tasks; it is beginning to rewrite the roles people play at work and in everyday life. As LLMs become collaborators, advisers, tutors, creators, and interfaces to knowledge, the value of many human roles is shifting from producing every output ourselves to directing AI systems, exercising judgment, validating results, and deciding what should not be delegated.
This talk explores the changes already visible in software teams, management, education, creative work, and daily decision-making—and what they suggest about the broader societal transformation ahead. It also examines the risks that come with convenience, including weakened critical thinking, blurred accountability, and growing dependence on systems we do not fully understand.
Rather than treating AI as simply another productivity tool, the talk presents it as the beginning of a wider redesign of work, institutions, and human identity. Attendees will leave with a practical framework for deciding what to delegate to AI, what should remain under human control, and which capabilities we need to strengthen as our roles continue to evolve.
Bio
Marcelo B. San Martin is a software engineer and QA automation leader with more than 13 years of experience, including work with Salesforce, PayPal, and Blizzard Entertainment. In recent years, he has focused on applied AI, LLM-powered systems, agentic workflows, and using AI to transform software engineering and quality practices. He also creates educational content exploring the AI revolution and its impact on work, creativity, critical thinking, and everyday life.
Leon Achata
TCS
RAG Is Not a Search Problem
Abstract
Modern RAG systems are often reduced to a simple pipeline: embed documents, perform vector search, retrieve the top-k chunks, and send them to an LLM. In practice, retrieval is rarely that simple. Exact identifiers, technical terminology, semantic paraphrases, lost document context, and multi-step questions can all expose weaknesses in a purely vector-based approach.
This talk explores how modern retrieval systems move beyond basic semantic search through hybrid retrieval, combining sparse and dense methods, contextual retrieval, reranking, query decomposition, and adaptive or agentic retrieval. Rather than presenting a single “best” architecture, we will examine the trade-offs between relevance, recall, latency, cost, and complexity, and look at when each technique is actually useful.
The central idea is simple: a RAG system does not need to find the most similar chunks; it needs to retrieve the right evidence for the question. We will build this idea step by step and show how retrieval quality, rather than simply choosing a more powerful LLM, can determine the quality of the final answer.
Bio
Leon Achata is a Generative AI Engineer focused on building production-grade LLM applications, RAG systems, and intelligent agents. He currently works on enterprise GenAI solutions at Apple via TCS, designing LLM-powered workflows, retrieval systems, and agent orchestration for large-scale international environments. Previously, he worked on multi-agent systems, RAG pipelines, and AI infrastructure across AWS and other cloud platforms. His background combines software engineering, AI/ML, and biomedical engineering, giving him a practical perspective on applying LLMs to complex real-world problems.
Josefina R. Casanova
IBM
From PDFs to Production RAG: Lessons from Building Document Pipelines with Docling
Abstract
Large Language Models have become remarkably capable, yet many AI projects still struggle to deliver reliable results. The problem is often not the model itself, but the quality of the information being provided to it.
In this session, we'll take a practical look at document intelligence for production AI systems. Through real-world examples and a live demo using open source tools, we'll compare different approaches for transforming complex documents into structured knowledge ready for retrieval-augmented generation (RAG).
Attendees will leave with practical techniques for improving retrieval quality, reducing hallucinations, and building more trustworthy AI systems. The session is tool-agnostic and highly practical, focused on lessons that can be immediately applied to real projects.
Bio
Josefina Casanova is a Data & AI Architect at IBM and a Computer Science Engineer passionate about using technology to solve real problems. She spends her days working at the intersection of AI, data and business, helping organizations build solutions that turn messy data into systems that work in production. A lifelong learner, she regularly speaks at tech events, hosts podcasts, and enjoys sharing lessons learned from building with emerging technologies.
Federico Novelli
Protect Group
The AI Security Paradox: Quantifying Risk in the Age of LLMs and Agents
Abstract
As organizations adopt LLMs and agentic systems, security teams are facing a new challenge: how do we turn emerging AI threats into concrete business risk that can actually inform security decisions?
Prompt injection, sensitive data exposure, excessive agent permissions, and unsanctioned use of AI tools can introduce risks that are difficult to evaluate using traditional security models. At the same time, organizations cannot respond to every emerging threat by continuously expanding their security
budgets.
This session explores how to think about AI-oriented cyber risk, from identifying
relevant threats and estimating their likelihood and business impact to evaluating whether a security control is worth its cost. We will examine how technical AI risks can be translated into structured risk conversations and financial terms, helping organizations distinguish between risks that require
immediate investment and those where simpler controls or accepted risk
may be more appropriate.
The goal is not to predict the future of AI security, but to provide a way to reason about security decisions as LLMs and agentic systems become part of the enterprise technology landscape.
Bio
Federico Novelli is the Head of Information Security & Technology Operations at Protect Group, where he oversees global security policy enforcement, technology infrastructure, and cross-functional operations. An Industrial Engineer currently finishing an MBA; Federico specializes in bridging the gap between technical cybersecurity operations, risk management, and high-level strategy. He is also a Professor at the Faculty of Engineering at Universidad de Montevideo and holds a Postgraduate degree in Project
Management, bringing a practical, multidisciplinary perspective to security, risk quantification, and organizational governance.
Gorka Ludlow
Karma Digital
Apps, not chats. Shipping macOS products that run AI models on-device
Abstract
Most people meet AI as a chat box. But the models can run as real software on your own machine, doing a job inside an app instead of answering prompts in a thread. This talk walks through shipping macOS products that do exactly that, with the models running entirely on-device: transcription and speaker diarization, screen understanding through an MCP server, voice cloning that reads text in your own voice, and an augmented-reality assistant. Running locally reduces per-inference cost, works offline, and can keep your data private, in exchange for fighting model size, memory, latency, and macOS packaging. We share the honest engineering behind products already in users' hands: which models we chose and why, what worked, what broke, and when local is the wrong call. Apps, not chats.
Bio
Gorka Ludlow is co-founder of Karma Digital, where he builds and ships macOS AI apps that run models on-device, and does AI consulting.
Carla Marcela Florida Roman
Bee Zero
Science at Scale: AWS Infrastructure for Reproducible Scientific Experimentation
Abstract
Validating a scientific hypothesis is not just a matter of ideas. It also requires infrastructure capable of running experiments in parallel while ensuring traceability, reproducibility, and cost efficiency. In this talk, we present how we designed an experimental platform on AWS to support real-world academic research at scale. The system enables thousands of experiments to run across multiple datasets, algorithms, and experimental conditions, using elastic compute with AWS Batch, structured storage in Amazon S3, a vector database to organize experimental results, and a centralized dashboard to explore and compare experimental behavior. The scientific use case behind this infrastructure is part of a research effort developed in collaboration with Eduardo Di Santi at the University of Colorado Boulder, exploring the relationship between the geometric structure of data and learning efficiency. The platform was designed to evaluate these types of hypotheses empirically at scale. This talk does not replace the formal academic presentation of the scientific findings, which are currently being reserved for publication. The session focuses on the architecture, design decisions, experimental automation, reproducibility, scalability, and lessons learned from using cloud infrastructure to support advanced scientific research. If you work in research, MLOps, data science, or cloud infrastructure, this talk presents a real-world case study of how AWS can transform a complex scientific question into a robust, scalable, and reproducible experimental platform.
Bio
Carla Florida is the CTO of BeeZero, an AWS Community Builder, and the creator of CarliCode, a community where she shares content about technology, programming, and artificial intelligence. Passionate about AI and Generative AI, she develops AI-powered solutions while helping make these technologies more accessible to people across Latin America. She is a former Platzi instructor and an active member of the tech community, combining her technical expertise with her passion for teaching, creating, and sharing knowledge.
Joaquin Bonifacino
Zyte
Long Running Agents: The Harness Is the Product
Abstract
Most agent demos run for thirty seconds. Production agents run for hours, and almost everything that breaks at that timescale has nothing to do with the model. Drawn from building an agentic data-QA engine running at 20B+ tokens per week in production.
Why long-running agents fail: context degradation, silent tool failure, unrecoverable state, cost drift, and why demos never catch them.
How harness design answers each one: compaction and memory, per-role model routing, sandboxed execution, checkpointing and recovery.
Where to start: what to build first moving an agent from demo to production, and what to instrument on day one.
Bio
AI/ML Engineer at Zyte, technical lead for LLMNL, an AI-powered data-QA engine running agentic workloads at 20B+ tokens per week. Previously led Promtior's Augmented Behaviors Lab (applied R&D in agent memory, voice and multimodal). Teaches Applied AI at ORT University, Montevideo.
Ben Sterne
Match My Email
Beyond Vibe Coding: How AI is Changing Development Cycles
Abstract
Just a few years ago, all new software was designed and coded by humans. Today, AI tools like Claude Code, Codex, and GitHub Copilot are reshaping how software is built from the ground up. This talk explores how adopting an AI-native product development cycle transformed our team's workflow and boosted productivity. We will examine the specific engineering patterns, automated verification loops, and team workflows required to move reliably beyond informal 'vibe coding' into production-grade AI development.
Bio
Ben Sterne is the Product Lead for AI initiatives at Match My Email, exploring applied AI in the CRM space. He is an early adopter of AI technologies and built the first AI prototypes with top product and design executives at Salesforce and Slack to guide their AI strategies. He comes from a product design and data science background, building upon successful products as well as building new features and products from scratch. Day-to-day, he is building AI tools, drawing on 10 years of experience in the tech space working with companies like Slack, Salesforce, Sonos, and Google.
Axel Labruna & Maika de Jesus Esteves
Nubiral
AgentOps: The future is here
Abstract
What if the next leap in efficiency in DevSecOps didn’t come from more tools, but from agents capable of reasoning, coordinating, and executing on our behalf? Agent-based artificial intelligence is no longer just a futuristic promise: it’s already changing the way we design, automate, and scale engineering workflows. In this talk, we’ll explore how AgentOps enables us to take AI agents from an interesting demo to real, governable, and impact-driven architectures. We’ll present agent-based architectures applied to DevSecOps scenarios, demonstrating how to design agents with clear responsibilities, explicit contracts, traceability, and collaboration mechanisms. We’ll also review design patterns for building more robust, reusable, and secure solutions. Through practical examples and demos based on real-world developments, we’ll see how these approaches can accelerate activities such as analysis, artifact generation, validation, code review, documentation, security, and pipeline automation, thereby reducing operational effort, delivery times, and costs. Attendees will gain a concrete understanding of what AgentOps is, how to start designing useful agent-based architectures, which patterns to apply, and how to turn this trend into a real advantage for technology, product, and business teams.
Bio
Axel Labruna is a Business Solution Architect at Nubiral, specializing in digital transformation, cloud architecture, DevOps, and AI-powered software development. He helps organizations turn technology into business value, with a focus on GitHub, agentic development, and AgentOps. A speaker at technology events across Latin America and an organizer of DevOpsDays Buenos Aires, Axel brings a practical perspective on adopting AI, empowering engineering teams, and connecting technical innovation with business strategy.
Maika de Jesus Esteves is a DevOps Engineer at Boreal IT Solutions in Buenos Aires, working daily with AWS, Azure, Kubernetes and CI/CD pipelines. She co-organizes DevOpsDays Buenos Aires, teaches DevOps at Mundos-e, and is active in the KCD Argentina and AWS community scene. Trained in telecommunications, she found her way into cloud and automation and now speaks about GenAI in real operations, including her work on the WIRU Assistant presented at Tech4Impact by Naranja X.
Marcos Martinez
Loopstudio
From 1 Hour to 10 Minutes: Automating Purchase Orders with AI
Abstract
A US enterprise food and beverage distributor received about 100 purchase orders a day by email, each client with its own PDF or Excel layout. Eight people spent their full workday reading them and typing them into the ERP, at about an hour per order.
We replaced that with an AI workflow built on the Anthropic API. Orders now take ten minutes including human review, with over 90% field-level accuracy, at a fraction of a cent per document. Those eight people now spend their time on client relationships and on improving the product.
The model was the easy part. About 80% of the effort went into analysis, architecture and specification, much of it data work: reconciling an ERP and a warehouse that disagreed on the same records, and writing down validation rules that only lived in people’s heads.
This talk covers what made the system trustworthy enough to run the business on. On the AI side, structured outputs with per-field confidence scores, human review for the exceptions, per-client evals that catch what an aggregate score hides, and a versioned ground-truth suite that turns every prompt or model change into a measured decision. Around it, the parts most demos skip, integrating with the company’s existing systems and a secure deployment built to live inside an enterprise environment. Attendees will leave with a blueprint for taking an AI workflow from demo to production.
Bio
Marcos Martínez is the CTO at LoopStudio and a Software Engineer (UCUDAL) who built the company's AI practice, establishing secure AI-native development as an engineering standard. A former backend lead in Mercado Libre’s shipping division, he currently architects high-impact solutions for US clients, leading technical strategy from proposal through production delivery. Marcos is an active OWASP Uruguay participant and has positioned LoopStudio as a Certified Partner in the Claude Partner Network.
Martin Brian
Marvik
The Harness Auction: What Makes an AI Agent Actually Work?
Abstract
What makes an AI agent effective: a better model, the right tools, specialized skills, more context, subagents, or a verification loop? In this interactive talk, every audience member receives a limited budget and bids in a live auction for the components needed to assemble an agent harness. Participants then lock in their chosen combination and watch it enter a controlled benchmark gauntlet measuring correctness, regressions, safe execution, speed, and cost until only the winning harness remains.Through the results, we will explore which components materially improve an agent, which merely add cost and complexity, and why the system surrounding an LLM can matter as much as the model itself.
Bio
Martin Brian is a Senior AI Engineer specializing in the design, deployment, and operation of agentic systems in demanding production environments. At Marvik, he works as part of the AI team for a global payments platform, building agents that automate end-to-end software-development workflows across the organization.
Lucia Hernandez
Metodica Digital
The User Is Part of the Harness: A Year Running an Agent That Adapts to Cognitive Load
Abstract
Most harness talks optimize for the task. This one optimizes for the human operating it. Two years ago, at 35, I was diagnosed autistic, and it explained something I had been working around for decades: my executive function is inconsistent. Some days I can hold twenty projects; some days I can't answer an email. Tools built for an average user kept assuming I was the same person every day. So I built Digitana, a harness on top of Claude that acts as an external executive function. It is not tuned to do more. It is tuned to load me less. This is a field report from almost 600 numbered working sessions, and how it works: - User state as a control signal. Every morning the harness computes a 0–100 Cognitive Load Index from the phase of my menstrual cycle (hormonal load shifts my energy and focus week by week), meetings, hours already spent with the agent, social load and day of the week. The zone rewrites the agent's operating rules: how many options it offers, how long it answers, and how much it decides without asking. - Memory you can diff, and a harness that learns from use. Sessions, minutes and operational memory live on disk in git, not in a vector DB or inside a provider. The model never changes, but the harness does: every correction becomes a written rule, every session closes with what was learned, and skills update their own instructions from how they were used. Nothing the agent "will remember" exists unless it's persisted. - One agent per project. Fourteen projects, each with its own project-manager agent and a single task board, so delegation is the default and the human only does what only she can do. - Letting an agent act in WhatsApp groups with third parties. Two permission levels, a mention-only trigger, and a fail-closed tool guard: the agent can run the project's scripts and commit locally, but never sends, publishes or deletes on anyone's behalf. The goal of this talk is to show a different kind of value AI can add: beyond making a company more productive, it can also take care of the humans behind it.
Bio
Lucía Hernandez Lettier is a Uruguayan founder who sold her first software at 15 and has worked in technology for more than 20 years. She is the founder and CEO of Metodica Digital, a board member of the Ibero-American Federation of Young Entrepreneurs (FIJE), and, through Luz Divergente, advocates for neurodivergent leadership in business. She is a Vital Voices Visionaries Fellow. For the past year she has designed and run Digitana, the AI harness she uses as an external executive function.
Martin Rojas
Cybersecurity Specialist & Auditor
Breaking the GenAI Data Pipeline: Architectural Security and AI-DSPM with OWASP
Abstract
Modern LLM integration frameworks and autonomous agents shatter traditional application security assumptions. By combining system mandates, external context, and user inputs into a single token window—a vulnerability known as the "Flat Namespace"—systems become inherently incapable of distinguishing code instructions from adversarial data payloads. Aimed at senior platform engineers and developers, this condensed 30-minute deep-dive bypasses basic prompt engineering to audit the raw underlying GenAI data infrastructure. Guided by the OWASP GenAI Data Security 2026 initiative, we will dissect real-world, high-impact vectors: semantic embedding inversion (reconstructing raw data from pure mathematical vectors), GPU memory cross-context KV cache bleeding, and non-human identity (NHI) escalation in autonomous workflows. Attendees will leave with a practical architectural reference blueprint to establish an effective AI-DSPM (Data Security Posture Management) strategy. We will cover automated token redaction, deterministic vector namespaces, and the essential 90-day checklist to validate data resilience before deployment. Key Takeaways: *The Flat Namespace Paradigm: Why traditional perimeter security and model-level guardrails fail to shield the underlying agent infrastructure. *Data Pipeline Exploitation: Technical analysis of vector store vulnerabilities, GPU session bleeding, and unmasked telemetry leaks. *The AI-DSPM Reference Architecture: A lightweight blueprint for layered peripheral data validation under open-source OWASP criteria.
Bio
Cybersecurity Specialist & Auditor | OWASP Buenos Aires Chapter Leader Professional with over 20 years of experience designing and auditing critical infrastructure and data security architectures for enterprise environments. He is a Cybersecurity Technician and is completing his Bachelor’s degree in Cybersecurity at the National University Raúl Scalabrini Ortiz (UNSO). Passionate about open-source tech communities, he is the Co-Founder of Nexus Cyber Academy and a member of the Internet Society (ISOC LATAM). He currently serves as the Leader of the OWASP Buenos Aires Chapter, actively driving the adoption of emerging security frameworks for modern GenAI engineering and cloud-native data pipelines.
Fede Rodrigo
CIO Office
AI Is Never the Starting Point
Abstract
Most AI initiatives begin with “Where can we use AI?”—before establishing what the organization wants to achieve, who is involved, and which constraints matter. Drawing on technology leadership experience and original research, this talk explores how that sequence can create complexity, dependencies, and lock-in even when the technology works successfully. Each successful fix can make the next change harder, gradually narrowing the organization’s options. Attendees will leave with a practical framework for evaluating AI initiatives by starting with people, context, and intended business outcomes, while considering the future choices their decisions may close off.
Bio
Fede Rodrigo is a Chief Information Officer (CIO) with more than 15 years of IT leadership experience across the US, Europe, and Latin America. He is the author of How to Become CIO and Succeed and holds an MBA from the University of Illinois Urbana-Champaign and a Computer Science degree from the Catholic University of Uruguay. His recent research on technological dependencies and lock-in informs his upcoming TEDx talk, “Technology Is Never the Starting Point.”
