Patents
2
1 pending · first 18 months
ActiveDublin, IrelandAI Engineer — Agents, Evals, Guardrails
Tale of the tape
Sev-1 record
2–0
resolved <12h and <4h
Eval conversation success
77–86%
from 30–40% · LLM user-simulator rebuild
CI pipeline
~30s
from 10–15 min · record/replay proxy
API spend removed
$35–70K/yr
LLM-call scoring replaced with local compute
Stack owned
runtime → evals → RAG
end to end
Main card
Main event
Agent Policy Layer
Python · protocol design · CI/CodeQL
Built an open-source guardrails protocol that enforces agent policies at the call site — runtime-agnostic, fail-closed by default, one line to instrument.
Co-main
Banco do Brasil — agent migration
Multi-agent systems · Tekton · production ownership
Led a squad migrating a bank's PIX, Cards, and Accounts agents, and resolved two production Sev-1s under load — one by re-architecting an execution graph, one with an idempotent single-writer registration toolkit.
Featured
The eval rebuild
LLM evaluation · proxy caching · patent-pending
Rebuilt an agent evaluation stack — an LLM user-simulator, a record/replay proxy, and a local semantic analyzer — moving conversation success from 30–40% to 77–86% while cutting run cost and time.
Prelims
- RAG migration off Milvus0 regressions
- agentevaluatoropen-source · Claude Skill
- PION GPU port · DIAS2.5× speedup
- Heart-sound segmentationsignal research · IISc
The corner
I take longer to learn things than most people. Two to four times longer. I've stopped apologizing for it, because what comes out the other side doesn't lose.
In my corner — coffee, iron, and the long way round.
Post-fight
Reach me directly.
- Résumé
- PDF ↗