> Hi, I'm Marcelo Prates!

I'm a senior AI/ML engineer and data scientist with 7+ years shipping production systems across GenAI, ML, and computer vision. I currently lead AI search at HopHR, where I build Moon — an AI-native candidate-search platform unifying retrieval, reranking, and explainable shortlisting into auditable hiring decisions.

Before HopHR I was promoted to Lead Data Scientist at Dataside, where I owned discovery, solution design, and delivery across multiple accounts. I built and shipped production agentic RAG systems (Azure OpenAI, LlamaIndex ReAct, pgvector), MLflow-tracked forecasting pipelines with Bayesian optimization, and uncertainty-aware evaluation with conformal prediction. I also led a team of three data scientists at ConstructIN, shipping four production computer-vision applications for 360° construction monitoring.

• HopHR (Moon) — Designed production AI search with Elasticsearch filters, semantic retrieval, multi-criteria reranking, and LLM-based assessments; built a PydanticAI multi-agent runtime with memory-aware context, SSE streaming, and observability for long-running workflows.
• Dataside — Agentic RAG with tool-augmented retrieval, stateful memory, and graceful degradation; hierarchical demand forecasting with conformal uncertainty; imbalance-aware classifiers; production ML on Databricks with PySpark, MLflow, and CI/CD retraining loops.
• ConstructIN — Led a team of 3 data scientists; semantic segmentation for 360° site monitoring; analytics dashboard for project tracking and decision-making.
• SingularityAI — Real-time AI voice assistant with ElevenLabs TTS, AssemblyAI STT, LangGraph agents with PostgreSQL checkpointing, and a YouTube → ChromaDB knowledge pipeline.
• Samsung R&D Institute Brazil — Led wearable-health ML from research through global Galaxy Watch rollout; memory-efficient inference for constrained devices; coordinated delivery with HQ.
• Condati — Rebuilt bid-optimization and forecasting for digital marketing; Prophet-based time-series with cross-validation, A/B testing, and automated model-renewal.
• Poatek (consulting), Vortigo — internal assistants grounded on proprietary code and spreadsheet sources, applied ML across optimization, NLP, computer vision, and risk modeling.

I hold a PhD in Computer Science from UFRGS (Aug 2015 – Jul 2019), focused on Geometric Deep Learning / Graph Neural Networks, the Ethics of Artificial Intelligence (especially machine bias), and Neural-Symbolic Reasoning.

My work has been published at AAAI 2019, IJCAI 2020/2021, and NCA 2020, with 1,400+ citations (h-index 8). I built custom GNN-based solvers for NP-hard combinatorial problems and AI bias measurement frameworks — the same measurement-and-evaluation muscle I now use in production.

My Erdős number is 3, through Moshe Vardi.

The tools I reach for, grouped by what I usually build with them.

GenAI & LLM Systems

OpenAI APIs, Azure OpenAI, LlamaIndex, LangGraph, PydanticAI, Agentic RAG, ReAct tool-using agents, DSPy, hybrid retrieval & reranking, structured outputs, Vercel AI SDK.

ML & Computer Vision

PyTorch, Scikit-learn, XGBoost / LightGBM / CatBoost, MLflow, Bayesian optimization, conformal prediction, segmentation, multi-GPU training.

Data & Infrastructure

PostgreSQL, pgvector, Redis, Elasticsearch, Supabase, Databricks, Snowflake, PySpark, Docker, Kubernetes, CI/CD, OpenTelemetry, Prometheus, Grafana.

Product Engineering

Next.js 15, React, TypeScript, Tailwind, FastAPI, Pydantic, SSE streaming, Vercel, Azure Container Apps, Playwright.

I've been making generative art and creative-coding sketches as a hobby since 2015. There's a small but growing gallery on theWork page (filter by tag: art).

My main interests are the intersection of art and the exact sciences, the nature of the artistic process in the context of generative art, and the relationship of generative art with different media — 2D printing, 3D printing, pen plotting, projections, interactive sketches. Other favorite playgrounds: physical and biological simulations, mathematical art, complex systems, signed distance functions, cartography, fractals, and abstract procedural art.

Themes I keep coming back to across work and side projects: autonomous AI agents and tool-using systems, uncertainty-aware evaluation (conformal prediction, Bayesian optimization), research-to-production pipelines, and the intersection of generative art with the exact sciences. I'd like to spend more time on neural-symbolic reasoning and on robust benchmarks for multi-step agent workflows.