Artificial intelligence talks and training
I share technical work on artificial intelligence from Uruguay. These proposals are for communities, universities, organizations and teams that want to understand how to design and evaluate AI systems using code and evidence.
The starting point is my public projects: architecture decisions, experiments, negative results and limitations. Each topic can become a talk about decision criteria or a workshop that examines an implementation and its results.
Proposed topics
AI agents that act: permissions, failures and recovery
Audience: Software and AI teams, developer communities and students with programming fundamentals.
What we explore
- Separate the model’s proposal from tool authorization and execution.
- Analyze a timeout without assuming the operation failed or repeating its effect.
- Examine traces, permissions and recovery tests in the reference projects.
Public reference material:
Evaluate models and agents before accepting an improvement
Audience: AI and data teams, quality specialists, and technology or product leaders.
What we explore
- Define a task, acceptance criteria and a versioned evaluation set.
- Distinguish task success, constraint compliance and regressions by segment.
- Interpret results, failures and cases requiring review instead of automatic acceptance.
Public reference material:
RAG with evidence: retrieval, languages and answers
Audience: Teams building document search or assistants, and engineering or data science students.
What we explore
- Separate retrieval quality from answer quality.
- Compare lexical, dense and hybrid approaches on a defined corpus and query set.
- Examine ranking errors and differences between English and Spanish queries and documents.
Public reference material:
AI inference: measure latency, memory and quality
Audience: Platform engineers, AI teams and people making infrastructure decisions.
What we explore
- Distinguish first-token wait, generation time and total duration.
- Measure weight storage and process memory as separate variables.
- Read a quantization experiment without extrapolating its results to another model or hardware.
Public reference material:
Formats to discuss
- Technical talk
- A presentation centered on a problem, the decisions that shape it and the available evidence, with time for questions.
- Code workshop
- A walkthrough of a public repository, its configuration and its tests. Scope depends on the group’s experience, time and available environment.
- Session for technology leaders
- A discussion of criteria for evaluating an AI proposal: what to test, which risks to control and which evidence to request before deciding.
How a session is prepared
- We agree on the audience, prior experience and the decision or skill the session should help develop.
- We select a topic and public materials that fit the scope. For a workshop, we define technical requirements and what will be run or reviewed.
- We distinguish observations from the experiments from claims that still need validation in another environment. References let participants continue working after the session.
Short introduction for organizers
Lucio Durán is C.T.O. at LOi and works from Uruguay. His public work combines software architecture with applied artificial intelligence, including open-source projects on agents, language model evaluation, RAG and inference. He publishes code, methods and results alongside each experiment’s scope and limitations.
Propose an invitation
Send the organization or event name, audience, topic of interest, tentative dates, in-person or remote format and available time. With that context, we can discuss scope, requirements and terms.