I’m Rafael Kovashikawa. I build LLM systems and the checks that keep their answers honest. At Fuse I work on agents and MCP servers for marketing analytics, and I built Emet, a pipeline that verifies what a model says against the data it was given. Before that I did macro and credit data science in Brazilian finance. I trained as an economist at FGV and live in Washington, DC.

I will use this space to post about things that I’ve found interesting along the journey and, of course, my projects.

What I do best

  • Data Science & Modeling · Develop and productionize time-series (ARIMA, VaR), tree-based and econometric models for CPI forecasts, credit risk and trading signals.
  • Data Engineering · Build and orchestrate ETL on Airflow + Kubernetes.
  • ML in Production · Deploy credit-risk and anomaly-detection models (FastAPI, Docker).
  • LLM Systems · MCP servers, agent orchestration, verification of model output, RAG pipelines.

Tech Stack

Programming: Python, TypeScript, JavaScript, SQL, R
LLM & Agents: MCP (Model Context Protocol), Pydantic AI, LangChain, context engineering, evals
Data & Cloud: Airflow, FastAPI, MongoDB, Redis, AWS, Google Cloud Platform, GitHub Actions
ML & Statistics: scikit-learn, XGBoost, LightGBM, Statsmodels, NLP
Languages: Native Portuguese; fluent English and Spanish

Professional Experience

Fuse (AI Data Strategy for Marketers), Data Scientist & AI Engineer
Jan 2026 - Present

  • Built Emet, a verification pipeline for LLM answers. It matches the numbers in an answer against the JSON returned by the tools that produced it, then runs semantic and repair agents on claims that cannot be matched numerically. Each answer gets a verdict of verified, partial or failed. In production since August 2026.
  • Architected the enterprise Salesforce integration as an outbound hosted-MCP client: OAuth 2.0 + PKCE authorization flow, cross-tenant refresh-token assertions, and an untrusted-content envelope guarding against prompt injection from CRM record text. Shipped as a distributable 2GP managed package.
  • Built 6 production MCP servers (Klaviyo, SEO, Salesforce, TikTok Ads, LinkedIn Ads, Exa) exposing 30+ analytics tools to LLM agents.
  • Designed a multi-agent ReAct orchestration layer for three research agents: an 8-action loop gated by a complexity classifier, adaptive retry across 9 error classes, and per-datasource circuit breakers.
  • Built a sandboxed executor for LLM-generated analysis code that intercepts hallucinated imports and redirects agents to the real API, so they self-repair instead of failing.
  • Engineered context system across 16 agents (per-API field references, deterministic tool-routing evals) to resolve vocabulary collisions across 12+ MCP servers and cut hallucinated queries.
  • Instrumented OpenTelemetry tracing and per-tool credit metering, turning LLM spend into a billable, auditable unit.

GYRA+, Head of Data
Aug 2024 - Jan 2025 · Brazil

  • Led the strategic re-architecture of data models and ETL/ELT pipelines (Python, Airflow, AWS), cutting AWS costs for the data team by 50%.
  • Built a high-performance simulation API (FastAPI, Polars, MongoDB) deployed on Kubernetes, enabling rapid credit-portfolio stress testing.
  • Delivered a production-ready RAG pipeline (LangChain, MongoDB) that extracts and structures key information from legal documents.

JGP Asset Management, Macro Data Scientist
Jan 2021 - Jul 2024 · Brazil

  • Designed LATAM CPI forecasting models and global macro dashboards that out-sped Bloomberg.
  • Engineered resilient pipelines (Python, SQL, Airflow) to ingest and curate macro datasets for trading research.
  • Developed NLP models to score Brazilian Central Bank communications and integrate sentiment into trading strategies.
  • Automated data workflows, eliminating manual steps and improving data timeliness.

Órama (acquired by BTG Pactual), Data Analytics
Sep 2019 - Aug 2020

  • Mined client data with Python and SQL, applying NLP and k-means clustering to improve investment recommendations.

IBM, Data Operations
Apr 2019 - Aug 2019

  • Automated ticket-management workflows with Python, Selenium, and SQL, eliminating +90% of manual effort

Education

MIT, Data Science & Statistics MicroMasters (edX)
May 2024 - Feb 2026

Fundação Getúlio Vargas (EPGE), B.Sc. Economics
2017 - 2021

Awards

Award What I Built
Google Cloud AI Hackathon (Dec 2025) Built MedAnnotator, a multimodal medical annotation system. [Post] [Code]
Big Data Hackathon (Sep 2021) XP Inc. & Microsoft Azure. Built a 14 GB big-data application on Azure.
Cryptocurrency Datathon (Sep 2020) FGV & Ripple. ML strategies for Bitcoin trading with NLP on news and Reddit.

Outside the terminal

I was born and raised in Rio de Janeiro and now live in Washington, DC. I like dogs, running, and Fluminense.

Benito at sunset Post-run selfie

Need a hand with AI and data problems?
Hit me on LinkedIn or open an issue and let’s chat.