Alberto Romero

Alberto Romero is the co-founder and chief executive of Jointly AI, where he builds autonomous agents for insurance, healthcare, and other regulated industries. He developed Meta-Adaptive Context Engineering, a framework that adjusts how agents reason, verify their work, retrieve information, and allocate computing resources.

Previously, Romero co-founded HumanAI and served as its chief technology officer, applying machine learning to mobility-risk prediction. He later led Citibank’s generative AI engineering team before founding Jointly.

In February 2026, Jointly introduced an autonomous insurance brokerage platform that coordinates specialized agents to collect customer requirements, contact insurers, compare quotes, and recommend policies. Jointly AI Broker combines confidence scoring, explicit workflow states, observable tool calls, and auditable decisions. Romero also developed Chloe, a healthcare voice agent, for medical preauthorization, coverage verification, multilingual conversations, and escalation to human staff.

• Task-aware agent optimization: Meta-ACE uses a learned controller to choose among concise prompts, additional reasoning, structured memory, verification, and selective model adaptation according to each task’s difficulty, uncertainty, and computational budget.
• Hierarchical verification: Model self-assessment, confidence-weighted agreement among different models, and executable checks such as API validation or schema compliance protect agents against unreliable feedback. Romero cautions that consensus still fails when multiple models share the same error.
• Accountable autonomy in regulated industries: Insurance and healthcare agents require domain-specific workflows, visible execution, calibrated confidence, and human intervention when warranted. Romero applies these safeguards to both insurance brokerage automation and medical administration.

Unresolved challenges include sparse training rewards, orchestration overhead, correlated verification failures, and the substantial data required to train adaptive controllers.