I build AI systems and publish the research behind them.

AI Engineer at Valor Capital Group. Computer Engineering at Unicamp.

Selected workRead the papers
Portrait of me

  • 1st of 102

    Computer Engineering at Unicamp, GPA 3.94/4.0

  • 2nd of ~1,000 teams

    Itaú Asset Quant AI Challenge 2025

  • 60 of 2,000+

    Selected for MBZUAI UGRIP 2025, then Best Team among 15 groups

  • 90 of ~105,000

    Selected for the Santander Open Academy at IE University

Graph of work

Every graph starts as text.

This is my profile, the same file I give to AI agents. First, pull out the entities.

Selected lines of llms.txt with the entity mentions highlighted.
  1. 1# Matheus Ferracciú Scatolin
  2. 5AI Engineer @ Valor Capital Group | Computer Engineering @ Unicamp (Ranked 1st of 102) | Fellow Estudar '25 | Published AI Researcher
  3. 26- 2nd place, Itaú Asset Quant AI Challenge 2025, among ~1,000 teams and 2,500+ participants
  4. 27- Best Team Award, MBZUAI UGRIP 2025 (selected in the top 3% of 2,000+ international applicants)
  5. 81- Day-to-day work spans agentic AI pipelines, knowledge graphs, and MCP-based tooling
  6. 83### Instituto Kunumi (kunuminst COLABS) | AI Researcher
  7. 86- Research on the automatic generation of Knowledge Graphs and new techniques for Graph-RAG systems
  8. 89### Enter | AI Fellow | Mar 2026 - Jun 2026
  9. 112### Semantix AI | Research Fellow / AI Researcher
  10. 114- Designed, developed, and led STELLAR

Then connect them.

Entities become nodes and sentences become relations. The two mentions of knowledge graphs resolve to one node.

Knowledge graph built from the mentionsFifteen nodes and sixteen relations. Both mentions of knowledge graphs resolve to one node. Matheus links to Valor Capital Group, Instituto Kunumi, Unicamp, Fundação Estudar, Itaú Asset Quant AI Challenge, MBZUAI UGRIP, Enter and STELLAR.engineer_atranked_1st_of_102placed_2ndresearched_atMatheusValor CapitalGroupValor CapitalGroupInstitutoKunumiInstitutoKunumiknowledge graphsknowledge graphsGraph-RAGGraph-RAGagentic AIpipelinesagentic AIpipelinesMCP-basedtoolingMCP-basedtoolingEnterItaú Asset QuantAI ChallengeMBZUAI UGRIPBest Team AwardSTELLARSemantix AIUnicampFundaçãoEstudar

What am I working on now?

Agentic AI pipelines, knowledge graphs and MCP-based tooling at Valor Capital Group, after Graph-RAG research at Instituto Kunumi (2025-2026).

What I am working on nowFrom Matheus to Valor Capital Group, then to agentic AI pipelines, knowledge graphs and MCP-based tooling. From knowledge graphs to Instituto Kunumi, then to Graph-RAG.engineer_atranked_1st_of_102placed_2ndresearched_atMatheusValor CapitalGroupValor CapitalGroupInstitutoKunumiInstitutoKunumiknowledge graphsknowledge graphsGraph-RAGGraph-RAGagentic AIpipelinesagentic AIpipelinesMCP-basedtoolingMCP-basedtoolingEnterItaú Asset QuantAI ChallengeMBZUAI UGRIPBest Team AwardSTELLARSemantix AIUnicampFundaçãoEstudar

Relations in the graph

  • Matheus, engineer at, Valor Capital Group
  • Matheus, ranked 1st of 102, Unicamp
  • Matheus, fellow of, Fundação Estudar
  • Matheus, placed 2nd, Itaú Asset Quant AI Challenge
  • Matheus, researched at, MBZUAI UGRIP
  • MBZUAI UGRIP, awarded, Best Team Award
  • Valor Capital Group, works on, agentic AI pipelines
  • Valor Capital Group, works on, knowledge graphs
  • Valor Capital Group, works on, MCP-based tooling
  • Matheus, researched at, Instituto Kunumi
  • Instituto Kunumi, generates, knowledge graphs
  • Instituto Kunumi, researches, Graph-RAG
  • Graph-RAG, retrieves over, knowledge graphs
  • Matheus, was fellow at, Enter
  • Matheus, led, STELLAR
  • STELLAR, built at, Semantix AI

Selected work. Four projects: two published, one in production, one placed 2nd of ~1,000 teams.

Itaú Asset Quant AI Challenge, 2025

KernelNet

A market-neutral trading strategy that replaces static correlations with nonlinear causality networks.

  • Sharpe 1.29
  • 54.85% annualized vs 22.78% benchmark
  • 2nd of ~1,000 teams

Read the case study: KernelNet

Semantix AI, 2024-2025

STELLAR

An LLM architecture for reliable customer support, built as a directed acyclic graph of nine specialized modules and eleven workflows.

  • First-author paper, JBCS 2026
  • Qualis A2
  • Scopus-indexed

Read the case study: STELLAR

Experience. Seven roles since 2024, in industry and research.

  1. Valor Capital Group

    Jun 2026 - Present

    Me at brunch in San Francisco with my boss, the other Tech Summer intern and a friend

    AI Engineer, Tech Summer · Industry

    Agentic workflows, data pipelines and technical diligence, working directly with José, the firm's Head of AI.

    • Agentic workflows
    • Data pipelines
    • Technical diligence

  2. Enter

    Mar - Jun 2026

    Me with my Enter AI Fellowship cohort in front of the Enter logo

    AI Fellow · Industry

    Turned a local script into a distributed production pipeline (FastAPI, Hatchet, React, LLMs) that processed thousands of judicial decisions.

    • FastAPI, Hatchet, React, LLMs
    • Featured on Enter's blog

  3. XP Inc.

    Jan - Feb 2026

    Me taking a selfie with Guilherme Benchimol, founder of XP Inc., in São Paulo

    Machine Learning Summer Intern · Industry

    Churn prediction MVP for high-net-worth clients with XGBoost, validated out-of-sample and out-of-time.

    • XGBoost
    • Out-of-sample and out-of-time

  4. Instituto Kunumi

    Aug 2025 - Aug 2026

    Me working on my laptop at a shared table

    AI Researcher · Research

    Automatic knowledge graph generation and Graph-RAG question answering.

    • Knowledge graphs
    • Graph-RAG
    • Entity and relation extraction

  5. MBZUAI

    Jun - Sep 2025

    Me wearing a Brazilian flag at the main entrance of MBZUAI in Abu Dhabi

    UGRIP Research Intern · Research

    3D brain tumor segmentation, missing-modality synthesis and response prediction, for the BraTS 2025 Challenge.

    • 3 papers
    • Best Team of 15
    • Top 3% of 2,000+

  6. Hyundai Motor Company

    Jan - Feb 2025

    Me with the other Hyundai summer interns in front of the Hyundai logo

    Data Analysis & ML Summer Intern · Industry

    Lead conversion model from 21% to 39% F1. Monthly data processing from 3 days to 3 minutes.

    • F1 21% → 39%
    • 3 days → 3 minutes

  7. Semantix AI

    Mar 2024 - Aug 2025

    STELLAR module diagram: a directed acyclic graph of nine specialized LLM modules

    Research Fellow / AI Researcher · Research

    Designed and led STELLAR. Built a hallucination benchmark of 7 models across 90,000+ questions.

    • STELLAR, JBCS 2026
    • 7 models
    • 90,000+ questions

Education.

I study Computer Engineering at , ranked 1st of 102 with a 3.94 GPA. I got into both Unicamp and USP straight from high school, after being valedictorian twice at and spending a Grade 12 semester in , with a 96% average. In 2026, Santander and picked 90 of about 105,000 applicants for two weeks in Madrid. I was one of them, and my team reached the Demo Day final.

  • Unicamp

    2023 - 2027 · Campinas

  • IE University

    2026 · Madrid

  • Lindsay Thurber

    2022 · Red Deer

  • Ilimit Educacional

    2016 - 2022 · Brazil

Research. Four papers, from LLM systems to medical imaging.

  1. Scatolin, M. F., & Pedrini, H. (2026).

    STELLAR: A Structured, Trustworthy, and Explainable LLM-Led Architecture for Reliable Customer Support

    Journal of the Brazilian Computer Society, 32(1), 128-144. Qualis A2, Scopus-indexed.

    First author
    DOI (opens in a new tab)Code (opens in a new tab)

  2. Tikhonov, D., Scatolin, M., Banerjee, M., Ji, Q., Jaheen, A., Salem, M., Elsayed, A., Wang, H., Hashmi, S., & Yaqub, M. (2025).

    Predicting Brain Tumor Response to Therapy using a Hybrid Deep Learning and Radiomics Approach

    BraTS-Lighthouse 2025 Challenge (MICCAI 2025). Submitted. arXiv:2509.06511.

    Co-first authorMean ROC AUC 0.81
    arXiv (opens in a new tab)

  3. Jaheen, A., Elsayed, A., Kim, D., Tikhonov, D., Scatolin, M., Banerjee, M., Ji, Q., Salem, M., Wang, H., Hashmi, S., & Yaqub, M. (2026).

    EMedNeXt: An Enhanced Brain Tumor Segmentation Framework for Sub-Saharan Africa Using MedNeXt V2 with Deep Supervision

    Lecture Notes in Computer Science, vol. 16376 (MICCAI 2025 / BraTS-Lighthouse), pp. 224-236. Springer, Cham.

    Co-authorLesionWise DSC 0.897
    DOI (opens in a new tab)Code (opens in a new tab)

  4. Banerjee, M., Ji, Q., Hashmi, S., Elsayed, A., Tikhonov, D., Scatolin, M. F., Jaheen, A., Kim, D., Wang, H., Salem, M., & Yaqub, M. (2026).

    MISFIT: Modality Inference via Style Fusion and Invertible Translation for Cross-Modality Synthesis of 3D MRI Volumes

    Lecture Notes in Computer Science, vol. 16377 (MICCAI 2025 / BraTS-Lighthouse), pp. 42-53. Springer, Cham.

    Co-author
    DOI (opens in a new tab)Code (opens in a new tab)

Profiles:Google Scholar (opens in a new tab)ResearchGate (opens in a new tab)ORCID (opens in a new tab)

Additional research: automatic knowledge graph generation and Graph-RAG at Instituto Kunumi (2025-2026).