Pedro Pestana

AI Product Owner at Ágora Investimentos. I specify and build systems with LLMs, agents and RAG, from proof of concept to production.

pmezherpestana@gmail.comLinkedInGitHub

About

In twelve months at Ágora Investimentos I went from intern to AI Product Owner. I have delivered 15+ proofs of concept and have 7+ systems in production, built with agents and RAG over market data.

I also wrote the team's internal RAG training, 63 lessons across 13 modules. Before that, at CSN, I automated back-office processes with AI and cut operational time by 80%. Business Administration at FGV-EAESP (2022–2026).

Experience

Aug 2026 – present

Ágora Investimentos

AI Product Owner

Promoted from Digital Products Intern, a role I held from Aug 2025 to Aug 2026

  • 15+ AI proofs of concept delivered and 7+ systems in production on the investment platform.
  • Multi-agent systems and RAG pipelines in production over market data, including the knowledge base the agents query.
  • Wrote and run the team's internal RAG training: 63 lessons across 13 modules, from chunking to operations, with a reference library built from scratch (BM25, vector indexes, quantization, reranking and evaluation).
  • Suitability classification models for fixed income.
  • Product vision, roadmap and backlog, and the model lifecycle alongside software engineering and data science.

Jan 2025 – Aug 2025

CSN

Energy Planning Intern

  • Back-office process automation with AI and traditional automation, cutting operational time by 80%. APIs, agents, OCR, Excel and VBA.

Jul 2023 – Jul 2024

Liga de Logística FGV

Head of Human Resources

  • Led 25+ members and restructured the selection process and performance KPIs.

Projects

  • CVM Analyzer

    Checks FRE, DFP, FCA and IPE filings against CVM Resolution 80: a 156-item checklist, a semantic diff between versions and a multi-agent RAG chat that cites the source passage. Academic exercise on public CVM data.

    CVM Analyzer landing page with the interface open on a company's Reference Forms: the list of versions and the PDF beside it.

    Python, FastAPI, RAG, Multi-agent

    Live demoCode

  • AI-generated image and video detector

    Nine deep learning models vote and a meta-learner decides. The answer comes with a confidence score and forensic analysis. FP16 inference with sequential model loading.

    Detector landing page with the result for one image: a 73% likelihood of being AI-generated and each model's score.

    PyTorch, ViT, CLIP, SigLIP, EfficientNet

    Live demoCode

  • Portfolio AI

    Portfolio management with a three-tier agent hierarchy over an ensemble of XGBoost, BiLSTM-Attention and Temporal Fusion Transformer, with live B3 data and walk-forward validation.

    PyTorch, XGBoost, FastAPI, Playwright

    Code

  • SLM Setup Wizard

    Fine-tunes a small language model on a single GPU, or on Colab. Ten steps, from scraping and dataset generation to LoRA/QLoRA training and GGUF inference with llama.cpp.

    SLM Local landing page with the wizard dashboard chatting with a model trained on Linux tech support.

    FastAPI, LoRA/QLoRA, llama.cpp, GGUF

    Live demoCode

  • CVM board network

    Who sits on the boards of Brazilian listed companies and how much power the overlaps concentrate. A Neo4j graph over public CVM data from 2021 to 2025, metrics with NetworKit and link prediction with Node2Vec and Random Forest.

    CVM board network landing page with the graph explorer: board members and officers as colored dots, some labeled with their names.

    Neo4j, NetworKit, Node2Vec, FastAPI, Sigma.js

    Live demoCode

  • Assay

    Personal finance with a multi-agent investment committee: you enter the contribution, and five chained stages research the market live and return an order sheet with ticker, quantity and limit price. It also tracks spending, cards and taxes. Runs fully local on a Raspberry Pi.

    Assay landing page with a completed R$ 5,000.00 contribution and the order sheet showing its first order, a daily-liquidity CDB.

    FastAPI, SQLite, Next.js, Multi-agent

    Live demoCode

  • RAG and MCP, explained

    Interactive material that shows what happens inside a RAG pipeline, from chunking to retrieval, and where MCP differs from a REST API. Runs entirely in the browser.

    RAG and MCP landing page with a 2D projection of the corpus chunks, grouped by topic.

    Next.js, TypeScript, Embeddings, MCP

    Live demoCode

  • LLM Visualizer

    The transformer architecture, one interactive step at a time.

    LLM Explorer landing page with the attention map for the sentence “o gato pulou no jardim” (the cat jumped in the garden).

    TypeScript, Next.js, Python

    Live demoCode

  • 8Patas

    Prototype of a veterinary platform that identified the pet's breed from a photo and built a care plan, a genetic-condition health profile and a preventive schedule from it. The page is now a showcase; the backend is no longer running.

    8Patas landing page with a photo of a dog and the prototype screens: sign-up, breed confirmation, health and schedule.

    React, Supabase, Deno, Vision

    Live demoCode

  • Smart Plant Controller

    Self-watering planter on an ESP32-C3 with a custom two-layer PCB, a capacitive soil probe and a pixel-art pet that reacts to the plant.

    Hoya Pet landing page with the plant and the pixel-art pet holding an umbrella, which is how it looks when the soil is soaked.

    ESP32-C3, KiCad, OpenSCAD, FastAPI, Supabase

    Live demoCode

Stack

Generative AI
LLMs (OpenAI, Claude, DeepSeek), Multi-agent systems, RAG and Agentic RAG, Embeddings and vector stores, Hybrid search and reranking, Structured outputs, MCP and tool use, LangChain, n8n, Fine-tuning LoRA/QLoRA, Local SLMs (llama.cpp, GGUF)
Machine learning and deep learning
PyTorch, XGBoost, scikit-learn, Optuna, ViT, CLIP, SigLIP, EfficientNet, LSTM and BiLSTM, Temporal Fusion Transformer, Node2Vec, Time series, Purged K-Fold, Walk-forward validation
Data, engineering and operations
SQL, Python (pandas, NumPy), PostgreSQL, Neo4j, MLflow, FastAPI, Docker, Git, systemd, Caddy and TLS, nftables, Tailscale, OpenCV and PaddleOCR, Power BI, Excel and VBA, Next.js and TypeScript
Business
Product management, Discovery and prioritization, Agile methods and squads, Product metrics, Financial markets, CVM compliance, Valuation

Business cases

TC & Pinheiro Neto2025

Valuation of a medical startup using DCF, WACC and CAPEX projection, with the legal deal structuring done in partnership with Pinheiro Neto Advogados.

P&G2024

Quantitative market share analysis and a repositioning proposal for the Vick VapoRub brand, based on consumer data and the competitive landscape.

SECOVI2024

Financial and commercial feasibility study for a real estate development, with NPV, IRR and payback.

TDS-FGV2024

AI immersion program with hands-on computer vision, robotics and autonomous driving projects in Python.

Education

Degree

FGV-EAESP

Business Administration

2022 – 2026

Machine learning and data

  • Machine Learning Professional Certificate (IBM, completed)
  • Deep Learning & Reinforcement Learning and Supervised/Unsupervised ML (IBM)
  • MLOps with MLflow, Hugging Face, SageMaker and Azure ML (Duke)
  • AI for Knowledge Workers (UC Davis)
  • AI on Jetson Nano (NVIDIA)

Financial markets and security

  • ANBIMA courses: Portfolio Management, Financial Risk, FIDC and Fixed Income Indices
  • M&A and Valuation for Startups (FGV-EAESP)
  • Ethical Hacker (Cisco, 2026)

Languages

  • Portuguese, native
  • English, advanced (EF SET C2, 78/100)
  • Spanish, intermediate

Contact

pmezherpestana@gmail.com

LinkedInGitHub

© 2026 Pedro Pestana