AI/ML Engineer & Healthcare Operations Leader, Miami, FL
I build production AI systems inside the healthcare operation I run. As Director of Operations & Innovation at a value-based care organization serving 14,000+ patients, I lead a full-risk Medicare panel of 4,000+ members (4.5+ Stars, 3 consecutive years) and personally design, ship, and maintain the ML models, LLM applications, and analytics platforms behind those results: a 34% panel-deficit turnaround, a 41% reduction in avoidable ER visits and admissions, and $50K+ in recovered revenue through Python/SQL automation.
- 🎓 M.S. Data Science & Artificial Intelligence, Florida International University (GPA 4.0/4.0)
- 📜 AWS Certified Machine Learning Engineer (Associate)
- 🌎 Trilingual: English, Portuguese, Spanish
- 🤖 LLM systems in live clinical workflows: an AI phone operator handling real patient calls end-to-end (3CX + RAG, human-in-the-loop escalation), and AI-assisted HCC risk-adjustment coding (LLM APIs, MCP integrations, LangChain)
- 📊 Value-based care analytics: HEDIS/Stars performance, HCC/RAF risk adjustment, and claims analytics processing ~150K claim lines per month
- 🏥 Utilization reduction that holds up: hospitalization-risk prediction feeding proactive CCM/TCM outreach, cutting avoidable readmissions 10 to 15%
- 👥 Leading 80+ clinical and administrative staff (10 providers, 13 departments) with SOP standardization and real-time KPI platforms
| Project | What it does | Stack | Outcome |
|---|---|---|---|
| HCC Coding Assistant | RAG + agent over 8,019 HCC-mapped ICD-10 codes: clinical language to validated codes, CMS-HCC V28 category, and RAF estimate, with all math in deterministic tools | LangChain 1.x, OpenAI, LangGraph memory | Every answer grounded in 2026 CMS reference data, zero codes from LLM memory |
| Readmission Risk API | FastAPI service predicting 30-day readmission risk for Medicare patients, with per-patient SHAP explanations | XGBoost, FastAPI, SHAP, Docker | AUROC 0.797 on 67 RFE-selected features |
| Medicare 30-Day Readmission (MIMIC-IV) | Reproducible readmission-prediction pipeline with a peer-review-grade no-leakage protocol | Python, gradient boosting, MIMIC-IV v3.1 | FIU capstone refactored for publication |
| Avoidable ED Visits Intervention | Data-driven intervention targeting avoidable emergency department visits in a Medicare Advantage population | NYU-EDA algorithm, Python, operational workflow design | 15,000+ ED claims analyzed, 40% avoidable, $2M+ in annual savings opportunities identified |
| VBC Financial Dashboard (private/production) | HIPAA-conscious value-based care platform: claims from 3 insurance plans, real-time EHR data via FHIR R4 over SMART Backend Services, anomaly detection with Critical/High/Medium severity scoring, and per-patient surplus/deficit analytics | React, TypeScript, Supabase/PostgreSQL, Cloudflare Zero Trust, CI/CD | Closes the 1-to-2-month visibility gap claims files leave; zero-trust with MFA, row-level security on PHI, and append-only audit logs |
| Codex HCC (private/production) | Multi-agent HCC extraction from medical records: a review agent surfaces candidate conditions, an independent validation agent verifies each against documented evidence | Python, Azure Document Intelligence, OpenAI, FastAPI | Precision +8.8pp across 11 tuning iterations at ~97% HCC recall |
| AI Phone Operator (private/production) | Voice agent fielding inbound patient calls: answers routine questions grounded in a curated clinic knowledge base, triages, and escalates clinical or sensitive calls to humans | 3CX, RAG over LLM APIs, knowledge base | Cut front-desk call volume and wait times with reliable after-hours coverage |
| KPI & Performance Tracking App (private/production) | Web app for SOP adherence, training reinforcement, and real-time staff performance tracking across clinical and administrative workflows | Web dashboards, KPI analytics | Drives targeted coaching and continuous improvement across 13 departments |
- Bandeira, T., Gonzalez, A., Poellabauer, C., Mondal, A.M. Predicting 30-Day Hospital Readmission in Medicare Patients: An Interpretable Gradient-Boosting Model on MIMIC-IV v3.1. M.S. Capstone, Florida International University.
- Bandeira, T. Using the NYU ED Algorithm and Admission-Hour Patterns to Reduce Avoidable ED Visits: A Primary-Care Perspective (2025). SSRN
Languages: Python · SQL · R · TypeScript ML: XGBoost · LightGBM · scikit-learn · SHAP · model calibration · feature engineering · survival/time-to-event analysis · NLP LLM & GenAI: LangChain · RAG with vector databases · agentic workflows · tool calling · MCP · prompt engineering · OpenAI & Anthropic Claude APIs Data: PostgreSQL · Supabase · pandas · Spark/PySpark · Delta Lake · Databricks · Power BI · Tableau Engineering: FastAPI · Flask · Docker · Git/GitHub · CI/CD (branch-per-environment) · React Cloud: AWS (ML certified) · Azure (Document Intelligence, OpenAI, Cognitive Services) · Cloudflare Healthcare: CMS-HCC V28 / RAF · HEDIS / Stars · HL7 & FHIR R4 / EHR interoperability · claims analytics · Medicare Advantage / full-risk VBC
- Agentic AI for clinical documentation and risk-adjustment review
- Real-time quality-gap closure systems that plug into care-team workflows
- ML that survives contact with production healthcare data: messy OCR, sparse labels, and all