Yuri Marca
AI Engineer & Solution Architect
Florianópolis, Brazil
Download CV LinkedIn GitHub Scholar Email
yurimarca@gmail.com · linkedin.com/in/yurimarca · github.com/yurimarca
I design, build, and ship production LLM systems end to end — from shaping the solution with clients in pre-sales through architecture, implementation, cloud deployment, and knowledge transfer. My focus is the layers that decide whether an LLM system works in production: evaluation, retrieval and data design, and compliance. I build evaluation harnesses that show where accuracy is lost, use them to drive architecture and roadmap decisions, and benchmark model, cost, and latency trade-offs up front.
I have worked remotely with US companies delivering agentic AI to clients in healthcare, travel, media, and real estate: HIPAA-compliant document pipelines, multimodal document automation integrated with ERPs, NL2SQL and RAG assistants, and real-time voice agents that handle calls and operate web applications. Before that, I spent years in industrial AI in Brazil taking ML from prototype to production — predictive maintenance and MLOps on AWS for wind turbines, and edge ML, computer vision, and IoT products. Earlier still, I did academic research on Bayesian optimization, Monte Carlo Tree Search, and multi-objective evolutionary algorithms (Best Student Paper Award, EMO 2019), and hold a Master’s in Engineering from Shinshu University as a MEXT Government Scholar.
I am looking to lead AI engineering and system architecture: setting technical direction, making trade-offs explicit, working closely with clients and stakeholders, and raising the engineering bar of the teams I work with.
Experience
Zazmic Inc
AI/ML Engineer
- Solution architecture & technical ownership — Technical owner of generative-AI engagements on GCP across healthcare, travel, and media, from pre-sales demos and estimates through architecture, delivery, and handover.
- Evaluation-driven engineering — Built evaluation harnesses, LLM-as-Judge suites, and trace-based root-cause analysis that pinpoint where accuracy is lost and steer architecture and roadmap decisions.
- Retrieval & data architecture for agents — Designed data models, retrieval strategies, and NL2SQL agents with abstention policies, and built RAG pipelines with incremental ingestion and citation-grounded answers.
- Regulated (HIPAA) AI systems — Architected event-driven document pipelines processing PHI (Document AI, Gemini, Pub/Sub, Cloud Run) with compliance by design, architecture decision records, and full monitoring.
- Real-time multimodal agents — Shipped voice agents that operate web applications through live native-audio models and function calling (Gemini Live, ADK, FastAPI WebSockets, React).
- Multi-agent pre-sales automation — Built a LangGraph multi-agent tool with human-in-the-loop review that turns discovery calls into cloud proposals, backlogs, and SOWs.
Dry Ground AI
AI Engineer
- Developed and deployed AI voice agents handling real estate calls, integrating tools to retrieve live information, send messages and email, and automate client interactions.
- Built a data science workflow for user profile classification and lead identification, supporting strategic sales initiatives.
WEG
Data Scientist
- Wind turbine anomaly detection — Deployed a production system on AWS SageMaker anticipating failures up to two weeks ahead and flagging ~70% of maintenance interventions, with a companion explainability API for root-cause diagnostics.
- End-to-end MLOps — Built a reproducible pipeline (MLflow, Docker, Kubernetes, AWS) covering automated training, versioning, deployment, and live Grafana/InfluxDB monitoring with full experiment lineage.
Macnica DHW
Data Scientist
- Smart sensors — Developed an SVM classification pipeline for vibration sensor data, applying NSGA-II to optimize the accuracy/compute trade-off for edge deployment.
- Computer vision APIs — Built a dual-layer video anonymization API (OpenCV + DBSCAN for motion-based person detection, ResNet-50-FPN as verification layer) tuned for low frame-rate surveillance footage.
- IoT development — Led end-to-end product development with full Docker containerization of the edge-to-cloud stack (MQTT, PostgreSQL, InfluxDB, Grafana), including on-device FFT computation.
Selected Work
Technical write-ups and code from projects I’ve built.
Classifying Time-Series Faults with PyTorch
Preprocessing multivariate vibration and acoustic sensor data from a machinery fault simulator, then training an MLP to classify faults — covering downsampling, rolling-window features, t-SNE visualization, and full evaluation metrics.
Building an End-to-End ML Pipeline
A reproducible pipeline for short-term rental price prediction, orchestrated with MLflow, configured with Hydra, and tracked in Weights & Biases — data ingestion through validation, training, and deployment.
Reproduction of the OCBA-MCTS paper — improving Monte Carlo Tree Search through Optimal Computing Budget Allocation, connecting directly to my ranking-and-selection research at Warwick.
More on GitHub, including music genre classification and image description with OpenAI.
Technical Expertise
Architecture & Practices Solution architecture ADRs Event-driven design HIPAA compliance
Languages Python C C++ Bash
LLM & Agentic Frameworks Google ADK LangGraph LangChain Gemini Live Vertex AI RAG Claude
Evaluation & Tracing LLM-as-Judge Langfuse LangSmith OpenTelemetry Grafana
ML & Deep Learning PyTorch Hugging Face scikit-learn XGBoost
Cloud GCP — Vertex AI, Document AI, Cloud Run, Pub/Sub AWS — SageMaker
MLOps & Deployment Docker Kubernetes MLflow Git GitHub Actions CI/CD
Full-Stack React TypeScript FastAPI WebSockets SSE streaming
Data Engineering PostgreSQL Cloud SQL Firestore BigQuery Redshift InfluxDB
Embedded ARM Cortex-M Jetson Nano Arduino FreeRTOS MQTT
Optimization Bayesian Optimization Multi-objective Evolutionary Algorithms
Certifications
Claude Certified Architect – Foundations
Anthropic
Research
University of Warwick, Warwick Business School
Doctoral Research
Ranking & selection, Bayesian optimization, and reinforcement learning. Investigated sample-efficient policy learning for constrained supply chain scheduling of biopharmaceutical products, developing Multi-Objective MCTS using marginal hypervolume as the selection criterion. Under Prof. Juergen Branke, collaborated with Prof. Chun-Hung Chen on two-stage ranking-and-selection, extending OCBA and Expected Value of Information algorithms.
Shinshu University
MSc Research
Developed a predictive model approximating complex decision space geometries caused by difficult Pareto set topologies, significantly improving MOEA solution distribution and convergence. Collaborated across the International Associated Laboratory LIA-MODO (Shinshu, Inria France, UAM Mexico), culminating in multiple publications and the Best Student Paper Award at EMO 2019.
Education
Shinshu University
MEng, Electronics and Information Systems
MEXT Monbukagakusho Scholarship (Japanese Government)
Concordia University
Exchange Student, Electrical Engineering
Science Without Borders Scholarship (Brazilian Government)
Federal University of Technology — Paraná
BEng, Electronics
Publications & Awards
🏆 Best Student Paper Award — EMO 2019 · Young Researcher Award — JPNSEC 2018 · Young Research Paper Award — IEICE Shin-etsu 2018
- Y. Marca, H. Aguirre, S. Zapotecas, A. Liefooghe, B. Derbel, S. Verel, K. Tanaka. Approximating Pareto set topology by cubic interpolation on bi-objective problems. EMO 2019, LNCS, Michigan, USA. (Best Student Paper Award)
- Y. Marca et al. MOEA with Cubic Interpolation on Bi-objective Problems with Difficult Pareto Set Topology. JPNSEC Journal, 2019. doi:10.11394/tjpnsec.10.12
- Y. Marca et al. NSGA-II with Spline Interpolation on Bi-objective Problems with Difficult Pareto Set Topology. JPNSEC Symposium on Evolutionary Computation, Fukuoka, 2018. (Young Research Award)
- Y. Marca et al. MOEAs on Problems with Difficult Pareto Set Topologies. IEICE Shin-etsu Branch IEEE Session, Niigata University, 2018. (Young Research Award)
- Y. Marca et al. Pareto dominance-based MOEAs on problems with difficult Pareto set topologies. GECCO ’18 Companion, ACM, 189–190. doi:10.1145/3205651.3205746
Full list on Google Scholar and Lattes.
Languages
Portuguese — Native English — Fluent Japanese — Conversational
Get in touch
The fastest way to reach me is email or LinkedIn. My full CV is available as a PDF.