Agents, retrieval and evaluation
Agents with scoped tools and permissions, RAG over owned sources and systematic evaluation of quality, cost and risk before any deployment.
AULAS LATAM’s lab: we explore emerging technologies with clear hypotheses, publish what we learn and only turn into products what gets validated.
Agents with scoped tools and permissions, RAG over owned sources and systematic evaluation of quality, cost and risk before any deployment.
Prompt injection, data leakage, tool abuse and red teaming, using the OWASP Top 10 for LLM Applications as the reference framework.
Cryptographic inventory, crypto-agility and testing with NIST standards: ML-KEM (FIPS 203), ML-DSA (FIPS 204) and SLH-DSA (FIPS 205).
Verifiable credentials for certifications, process traceability and tamper-evident records, always checking whether a conventional database is enough.
On-device AI, WebGPU, open data and new interfaces. We test early to decide with evidence, not hype.
Every experiment starts with a question. If the answer is “no”, that gets published too.
Classify and extract document data with per-field traceability and a human review queue.
Meaning-based search, in Spanish and English, over the Lexia glossary (Glosario IA by Tu Aula Latam).
Reproducible attack cases to evaluate agents that read email, documents and web pages.
Detect quantum-vulnerable algorithms in repositories and configurations.
Issue “Certificación IA para el Trabajo” certificates as credentials third parties can verify.
Connector to SECOP II open data to analyze technology procurement in Colombia.
Adjust pace and difficulty of short lessons from each learner’s answers on TuAula.
Small models running on-device with WebGPU for features that keep data off servers.
No experiments in this line yet.
An experiment ends when it answers its question, not when time runs out. It may become a product on aulas.lat, a capability on aulas.si, a course on tuaula.lat or a report on what we learned.
A measurable question and the criterion that would answer it.
The minimum needed to get evidence, with test data and no real personal data.
Metrics defined upfront, with risks and limitations documented.
Hypothesis, method, results and limitations, even when the result is negative.
Researchers, graduate students and specialists can propose experiments or join one in progress.