AI Engineering · Multi-Agent Systems
Senior software engineer and PhD researcher focused on agent orchestration, workflow design and agent evaluation — the reliability layer that turns LLM demos into systems teams can actually ship.
Three connected problems: making agents work together, proving they work, and keeping humans in control of what matters.
Designing multi-agent topologies — supervisor patterns, hub-and-spoke architectures, per-node model routing with cost circuit breakers — using LangGraph and Pydantic AI, with MCP for shared tooling.
Agent evaluation is the bottleneck nobody budgets for. I work on pass^k metrics, Reliability@Cost, failure taxonomies and verification gates — treating evals as CI, not as an afterthought.
Instrumenting agent workflows with OpenTelemetry GenAI conventions, and designing explicit human-in-the-loop gates so autonomy stays proportional to demonstrated reliability.
Independent research and engineering projects. Client work at Critical Software (automotive, safety-critical domains) is under NDA — happy to discuss the approach in conversation.
An agent evaluation and self-improvement framework: pass^k reliability metrics, Reliability@Cost, MAST failure taxonomy, Wilson confidence intervals and a honesty gate that blocks unverifiable success claims. Positioned as a reliability-science layer that couples to LangGraph or Temporal through OpenTelemetry GenAI semantic conventions.
A multi-agent architecture for cybersecurity incident response, combining LangGraph orchestration with LightRAG retrieval. Accepted at CENTERIS 2026 — agents triage, correlate and draft response actions under explicit human approval gates.
A library of transferable reasoning skills for coding agents — plan-then-execute, verification gates, debugging methodology, stop-and-escalate, context discipline and failure-mode awareness — grounded in empirical research on what actually transfers between models via text instructions.
A three-tier hub-and-spoke reference design for running fleets of coding agents: a single source of truth in AGENTS.md with per-agent shims, MCP for shared tools, and CI gates as the real standardisation mechanism rather than convention documents.
Before specialising in AI systems, I designed and shipped several tourism websites and platforms for European markets — multilingual, content-heavy products where performance, SEO and editorial workflows decide whether the site earns bookings. That delivery experience is why my AI work stays anchored to what survives contact with real users.
Doctoral Researcher in Data Science and Web Technologies (UTAD / Universidade Aberta), with ~8 years of industry experience. Senior engineer at Critical Software, working on large-scale automotive software. My research programme sits at the intersection of multi-agent systems, retrieval architectures and agent evaluation — the thesis is simple: autonomy without measurement is a liability, so the interesting engineering problem is building the measurement layer.
Degrees and certifications, verifiable by ID. The testing certification is not an accident — evaluation of AI systems is software testing taken seriously.
Tools chosen for one property: they let me reason about what the system is doing.
Open to conversations about agent orchestration, evaluation infrastructure and applied-AI consulting — from recruiters, and from teams whose agent systems work in the demo and fail in production.