ARTICLES
AI Engineering, software, and product decisions with context, trade-offs, and evidence. Articles preserve the record of each experiment.
AI Engineering · Machine Learning
In DefectRisk, more complex models delivered small gains. Group-aware evaluation, conflicting labels and calibration showed where to invest next.
AI Engineering · Systems
The decisions behind OpsPilot AI: evidence-grounded retrieval, deterministic authorization, approval bound to the exact action, and reconciliation of ambiguous external writes.
Applied AI · ML Systems
On the same held-out set, Laya went from 8/70 to 60/70 correct tuples after adaptation. An experiment in structured decisions, confidence, and the cost of control.
Applied AI · Engineering
A frozen synthetic held-out set compared deterministic rules, Ollama, and Jev 1.13. See the metrics, latency, cost, and limits of the experiment.
Security · Engineering
A Salus security review highlighted why authentication alone cannot authorize access to each patient and how to make ownership testable.
Applied AI · Engineering
How I used a deterministic baseline, a labeled dataset, and objective metrics to evaluate an LLM-based triage system instead of trusting subjective impressions.
Applied AI · Engineering
An operational triage benchmark showed where a local LLM beat deterministic rules — and where it didn't. The result led to a hybrid architecture with human review and an audit trail.
Engineering · Decisions
The architecture decisions behind this portfolio — and the evidence behind them — from the person who built it.
LET’S TALK
Professional opportunities, partnerships, or a conversation about AI Engineering, software, and product.
contato@marcelotaparelli.com.brConnect on LinkedIn