AI Security Architect · Secure Agentic Systems · Cloud Architecture
Portfolio · LinkedIn · Kaggle · Email
I work at the intersection of AI security, security architecture and AI engineering. My focus is what happens when untrusted content reaches an LLM system's context, memory, tools, APIs or code-execution layer—and how to turn those failure modes into threat models, reproducible tests and enforceable controls.
Current role: AI Security Architect at BBVA Technology
Public professional record: 200+ AI attack vectors assessed and seven critical findings in banking contexts. Professional context →
Here I publish practical labs, adversarial scenarios, agent harnesses and architecture prototypes that can be inspected and reproduced. Experimental, simulated and in-progress work is labelled explicitly.
- AI and LLM security — prompt injection, unsafe tool use, RAG and memory poisoning, adversarial evaluation and threat modelling.
- Secure agentic systems — tool and MCP boundaries, least privilege, execution isolation and context management.
- Cloud security architecture — AWS, IAM, containers, CI/CD, observability and secure-by-design delivery.
- AI engineering — Python, PyTorch, LangGraph, FastAPI, TypeScript and systems fundamentals in C.
A reproducible local lab for turning AI attack paths into deterministic tests and measurable controls.
threat modelling attack scenarios guardrail evaluation synthetic data
Scope: security prototype, not presented as production-ready. Open the deterministic demo →
A multi-runtime agent-harness template that makes skills, hooks, policies and CI-oriented controls inspectable.
agent harnesses runtime policy hooks CI controls
Scope: reusable engineering scaffold; effective security still depends on the host runtime and permissions.
AI engineering foundations built from first principles, beginning with tensors, autodiff and explicit evidence gates.
autodiff transformer foundations evaluation gates
Scope: active learning and engineering lab; the current implementation state is documented in the repository.
Attack-surface reconnaissance combined with Neo4j graph modelling and LangGraph-assisted analysis.
authorized reconnaissance Neo4j LangGraph
Scope: security learning project for explicitly authorized environments only.
A prototype combining multi-agent fraud analysis, RAG workflows and adversarial-testing components.
multi-agent systems RAG AI red teaming
Scope: portfolio prototype; not described here as a deployed production platform.
- Threat-model the system before choosing a guardrail.
- Treat prompts as guidance; enforce security through permissions, isolation and runtime checks.
- Prefer synthetic or explicitly authorized data for public experiments.
- Separate implemented, tested and planned capabilities.
- Publish enough evidence for another engineer to reproduce the result.
I am deepening three connected tracks: adversarial security for tool-using AI agents, cloud and IAM architecture, and the low-level programming foundations needed to reason precisely about runtimes and failure modes.
If you are working on secure AI systems, agent security or cloud architecture, the best starting points are GenAI Seguro Lab and ARCA Agent.