DEUS AI Guild
The talks we give each other.
19 sessions from 14 people on the team, on agents, memory, guardrails, observability and the tooling around them. Each has its own page with what was covered and what the room made of it.
Next session
Wednesday 11 November
RTK AI
Rodrigo Figueiredo
Archive
Every session, newest first
2026
-
XIX
28 October 2026 · Pablo Pousada
Real-time knowledge graphs for AI agents with Graphiti
Episodic memory for agents on a knowledge graph that keeps history: how Graphiti works, a local demo, and where it stops being a good fit.
-
XVIII
14 October 2026 · Bruno Cabado
LLM vs Jev (and openjev) vs Laya vs kev
Decision models answer a question in one pass with a probability per option. What they are, how they compare with an LLM, and how to test them on your own data.
-
XVII
30 September 2026 · Julian Silveira
Codegraph and other tools to turn a codebase into a graph
How code-graph tools work, and a small experiment testing the vendor's claims on a toy project.
-
XVI
19 August 2026 · Felipe Vieira
AI caching strategies
Three layers of caching for LLM applications, which one to reach for first, and where semantic caching goes wrong.
-
XV
22 July 2026 · Ana Maria
Neuro-symbolic AI
Pairing an LLM's creativity with a solver's guarantees, drawn from a summer school on neuro-symbolic AI.
-
XIV
8 July 2026 · Pedro Costal
Deep Agents (LangChain)
An agent runtime built on LangChain Core, and an honest answer on when a plain ReAct agent is still the better choice.
-
XIII
10 June 2026 · Iago Arán
Skills, agents and everything else
Why coding agents need skills, how orchestrators and sub-agents keep context small, and what separates a skill from an agent.
-
XII
27 May 2026 · Cris Frojan
How important is a golden dataset for LLM evaluation?
A small, human-verified reference set turns subjective judgements of LLM quality into numbers you can track.
-
XI
-
X
15 April 2026 · Pablo Pan
Proxmox walkthrough
An open-source virtualisation platform for fast, cheap test environments, and a plan to put it on idle office hardware.
-
IX
1 April 2026 · Julian Saborido
AI guardrails: strategies and options
A deterministic control layer around a probabilistic model, shown live against five scenarios.
-
VIII
18 March 2026 · Felipe Vieira
OpenCode
A terminal coding agent with plan and build modes, any model provider, and shared project configuration.
-
VII
4 March 2026 · Gilberto Plaza
PageIndex
Retrieval without vectors: an LLM reasons over a document's structure, shown on a contract-extraction use case.
-
VI
18 February 2026 · Diogo Mariano
Long-term memory
Transformers are stateless by design, so how do agents remember? Four kinds of memory, RAG, Letta and the open problems.
-
V
4 February 2026 · Julian Silveira
Exploring Docker Model Runner
Running and testing models locally with Docker's own tooling, and when to pick it over Ollama or vLLM.
-
IV
21 January 2026 · Rodrigo Figueiredo
Types of agent architectures
Reactive, ReAct, planning and hierarchical agents, and why architecture matters more than the model.
-
III
7 January 2026 · Pablo Pousada
Langfuse + LGTM stack for logs, metrics and traces
Why LLM applications need their own kind of observability, and how Langfuse and the LGTM stack split the work.
2025
-
II
11 December 2025 · Fernando Souto
SOLID, clean code and clean architecture
Three levels of clean code: daily practices, the SOLID principles and layered architecture.
-
I
26 November 2025 · Ana Maria
Universal Tool Calling Protocol (UTCP)
A lighter alternative to MCP: hand the agent a manual for each tool and let it call the tool's own API.
Guides
The stack behind the talks
- Python uv, ruff and ty: the tooling we build on.
- LangChain LLM chains that survive production.
- LangGraph Agent workflows with state, persistence and a human in the loop.
- Azure AI Search Vector and hybrid search, and RAG retrieval tuning.
- PostgreSQL Checkpointing and state management.
- Azure Container Apps Deploying and scaling agents on Azure.
- Langfuse Tracing, golden datasets and continuous evaluation.
- CI/CD Pipelines to build, test and ship LLM applications.