Researching how AI systems fail, adapt, and exceed their intended boundaries.

Pulled scheduling and patient-flow data, identified intake as the throughput constraint, redesigned the workflow into a digital pre-arrival system. Wait time down 50 percent, no-show rate down 40 percent, schedule density restored. Patient-ID reconciliation across Excel and the Peruvian double-surname system was the hidden engineering before any throughput analysis ran.

Diagnosed cross-system breaks with SQL, built per-batch validation checks at every join point, translated findings into a single Tableau view designed for leadership reading. Mismatches down 25 percent, manual reporting effort down 40 percent. Breaks now caught at ingest instead of weeks later.

Pre-visit triage from a sentence and an optional photo. Plain language in, schema-validated structure out: body systems involved, ranked possible causes, red flags, and the questions a clinician is likely to ask. v1 was a LoRA fine-tune of Qwen2.5-1.5B. v2 is multi-modal Claude Sonnet 4.6 vision. Same schema across both, so hallucination containment stays structural rather than dependent on training accuracy.

Population-health outreach prioritization for primary-care teams on HEDIS or value-based-care contracts. A synthetic 1,000-patient panel, USPSTF and HEDIS rule application, and a three-component prioritization scorer (clinical urgency, response likelihood, equity priority). Personalized outreach drafted with Claude using prompt caching, so per-message cost stays roughly flat across hundreds of drafts.

A runtime execution control layer for autonomous AI agents. Sentra sits between agent decision-making and tool execution. Every proposed action is evaluated against deterministic policy rules, cumulative risk is tracked across a session, and a three-strike rule shuts down agents that drift before they cause harm. Model-agnostic by design. Cross-industry: claims processing, customer communications, internal tooling, developer agents, clinical decision support.

Dual-LLM verification. One model runs as a worker, another stress-tests its output against user-defined rules before acceptance. Failures return as structured feedback so the worker can revise; persistent failures are blocked rather than silently passed downstream. Three-strike shutdown spawns a replacement agent with memory of why the last one failed.