Welcome to Leonardo Kuster's portfolio
/help for commands · cwd: ~/leonardo-kuster
@leonardokuster21 · just now
> ok. build my portfolio. make it perfect. zero bugs. zero mistakes. production-ready, obviously. waves are good today. i'm going surfing. you're the senior engineer now. it's all in your hands.
On it. Starting with his name.
Leonardo Kuster
AI Engineer. I build AI agents that reach production because they start with the data.
Data Engineer turned AI Engineer · Economist · Vila Velha, Brazil
press ⏎, scroll, or type /help
Bash(cat leonardo.jpg | ascii)
whoami
The human behind the prompt.
Leo is an AI Engineer at Falconi, a management consulting firm, on the Tech Consulting team. He's the team's go-to person on AI agents: he designs multi-agent systems and data pipelines with a human in the loop.
Before agents, he was a data engineer. He's been automating repetitive work since 2022, and with agents he got to 100% of his own tasks. Including, it seems, this portfolio.
Economist by training. Surfs. Plays guitar. Trains jiu-jitsu. Right now he's probably doing one of those.
- role:
- AI Engineer
- at:
- Falconi · Tech Consulting
- since:
- May 2025
- base:
- Vila Velha, Brazil
- degree:
- Economics · Fucape
- speaks:
- Portuguese, English
- off_the_clock:
- surfing, guitar, jiu‑jitsu
- status:
- …
Agent(agent-flow)
run flow
From backlog to pull request, with no one typing.
As a Data Engineer at Falconi, the volume of requests pushed Leo into AI. Agents ended up doing all of his work: pipelines, production bug fixes, Power BI reports. The clearest example is a fully autonomous agent flow, from backlog task to pull request, built with Claude Code and LangGraph. Pick a bug and let it run.
- picks up a task from the backlog
- analyzes the bug
- fixes the code in an isolated git worktree
- tests it in dev
- opens the pull request
- comments with evidence that the fix works
The evidence step is the point: the pull request arrives with proof that the fix works, not just a diff.
Waiting for a bug. Pick one above and press Run agent.
★ Task closed ★
To Do → In review
Backlog task to pull request with evidence, and nobody typed a line of code.
Bash(git log --graph --all)
git log
From spreadsheets to agents.
Every chapter is a commit, starting before the first job. Click one to see the diff.
cf704ff(HEAD -> main, tag: ai-engineer)feat: promoted to AI EngineerApr 2026
Falconi · Tech Consulting · São Paulo
- Lead the implementation of AI agents on the team as the subject-matter expert on agentic AI.
- Design and orchestrate multi-agent systems with LangGraph and LangChain for consulting projects.
- Build AI agents on OpenAI and Anthropic models, with tool use (function calling) and multi-step execution loops.
- Apply RAG, MCP (Model Context Protocol) and evals in agent projects.
- Build automated data pipelines with human-in-the-loop (HITL) review steps.
8312a93merge: branch 'economics' (Bachelor's in Economics)Dec 2025
Fucape Business School · 2021–2025
- Graduated in Economics from Fucape Business School.
- Econometrics and statistics gave me the habit of measuring business value, not just model accuracy.
9f21b24feat(agents): automate 100% of my own tasks2025–2026
Falconi · Data Engineer
- Facing a high volume of requests, automated 100% of my own tasks with AI agents: pipeline development and optimization, production bug fixes and Power BI reporting.
- Built a fully autonomous agent workflow, from backlog task to a tested pull request with evidence.
52d3b2efeat: join Falconi as Data EngineerMay 2025
Falconi · management consulting
- Built end-to-end data pipelines: multi-source extraction in Python into Azure Blob Storage, a medallion architecture on Databricks and Spark, modeling with dbt and orchestration with Databricks Workflows.
- Refactored a client's data engineering system to cut data ingestion and processing time and, with it, cloud costs.
- Built and optimized Power BI reports.
46c2ef7feat: ship the first data warehouse on Azure SynapseMar 2024
Apex Partners · Mid-Level Data Engineer · Vitória
- Served as the company's main data reference, responsible for every data integration and pipeline.
- Led the development of the data architecture and managed the data team, making technical and strategic decisions.
- Completed the company's first data warehouse on Azure Synapse, moving data out of xlsx and csv files on SharePoint and pipelines out of Python notebooks.
- Integrated the Sankhya ERP and the Salesforce CRM, making data secure and accessible across the company.
- Delivered visual management reports on top of the data warehouse.
33c6efcfeat: start the company's first data warehouseMar 2023
Apex Partners · Junior Data Engineer
- Took on multiple roles during a critical period for the tech team.
- Started building the company's first data warehouse on Azure Synapse, including its data orchestration.
- Built ETL and ELT pipelines on Azure with Python and Spark.
- Integrated data with Salesforce.
- Helped build dashboards at the strategic, tactical and operational levels.
eaeb560feat: first cloud architectureNov 2022
Apex Partners · Data Engineer Intern
- Helped design and implement the first version of the company's cloud architecture, including the first version of its database, built with Azure Functions.
- Maintained cloud data pipelines on Azure.
f78ac5a(quant-dream (never merged))revert: "become a quant"Jul 2022
The day the quant plan died
- Joined APX Invest, part of Apex Partners, the largest investment holding in Espírito Santo, where I was born and raised.
- No quant desk, no tech team, everything rudimentary. The quant plan died right there.
- Branch left unmerged. Data engineering was born instead.
276a8fc(branch: main)feat: first-week automation ends the interns' manual workJul 2022
APX Invest (Apex Advisory) · Data Analyst · Vitória
- Joined as a data analyst and helped build the data infrastructure from scratch.
- In my first week, built an automation that eliminated the manual work of a team of interns.
- Automated routines with Python (RPA) and built ETL pipelines and Power BI dashboards with operational data for investment advisors.
- Quickly promoted to the data engineering team at the holding company, Apex Partners.
c1072acfeat: join FGV Quant in São PauloMar 2022
FGV Quant · member, 2022–2023
- The only member from outside FGV.
- Developed quantitative models to forecast volatility and returns.
- The dream: quant trader or researcher.
5f2e1c7feat(ml): predict if an asset goes up or down in 5 days2021
Self-taught
- Dove into data science and machine learning: linear and logistic regression, decision trees, random forest, k-means, DBSCAN, neural networks, XGBoost, ensembles.
- One question behind all of it: up or down, five days from now?
d3a90be(branch: economics)init: Economics @ Fucape Business SchoolJan 2021
Fucape Business School · Bachelor’s in Economics
- Econometrics, statistics, finance, plus a seat in the college’s financial markets league. The habit of asking what something is worth before building it.
7b28230feat: learn Python to backtest every technique2020
Self-taught, out of pure annoyance
- Nobody showed historical performance data for their strategies, so I learned Python to test them against the past.
- First lesson, still valid today: data before opinions.
76662eefeat: day trade mini index and mini dollar on B32019
B3, the Brazilian stock exchange
- Learned a lot of techniques.
- Each one came with plenty of confidence and zero historical data. (See next commit.)
cfaca2a(branch: quant-dream)init: get rich on the internet2019
The internet, unfortunately
- Found binary options. Found out it was a scam.
- Lost the illusion, kept the market: fell in love with the financial markets.
Read(CLAUDE.md)
cat CLAUDE.md
How he thinks about agents.
Every agent Leo builds starts from the same four rules. Click a line to see why it’s there.
# How Leo builds agents
> conti, how much did we sell last month?
Confidently: “a lot.” Three files called final_v2 disagree, so I went with the one with the nicest colors.
Here’s the figure, the query that produced it and the table it came from, checked against last month’s close.
conti wants to run this on prod:
DELETE FROM sales WHERE created_at < '2025-01-01';Do you want to proceed?
fix: stuff
LGTM 🤷 trust me
- Tests added and passing
- Before and after query
- Dev run log attached
First we set up seven agents, a vector database, a router, a critic and a supervisor for the router. bla bla bla bla…
What does this process cost the business today, where does it break, and who signs off when the agent is wrong?
Bash(stack --tree)
my stack
Agents stand on data.
This tower is my stack: the tools I work with every day, from the data at the bottom to the agents at the top. Most AI agents stall at the pilot stage, and rarely because of the model: the problem is the data and the process around them. Pull a block and see.
Leo’s stack, from the ground up
- data: Azure, Azure DevOps
- data: Databricks, dbt
- data: Python, SQL, Spark
- context: Context engineering
- context: RAG, MCP, Evals
- agents: Claude Code
- agents: LangGraph, LangChain
- agents: CrewAI, AutoGen, Maestri
Tap a block to pull it.
WebFetch(fora-do-radar)
fora do radar
Under the radar.
Fora do Radar (“off the radar”) is Leo’s weekly series: reflections on AI, how he thinks about agents and curiosities few people have seen, plus really interesting repos with skills, agents and plugins. In Portuguese, every Wednesday on LinkedIn. The first edition is on its way, so the radar is still quiet.
Follow on LinkedIn ↗Bash(logout)
off the clock
Waves are good today.
While the agents work, Leo skimboards, plays guitar and trains jiu-jitsu. Try all three.
Part 1: hold Space, Enter or the action button (or hold on the picture) to run, and let go right after the front wave breaks to throw the board on its foam. Part 2: surf the wave behind it. Keep carving up and down the face with the mouse (or a finger, or ↑↓) to build speed; stop and you slow down and the wave catches up. Tricks: S goes upside down, D spins the board (or the big buttons). Put the mouse behind the board (or ←) to brake into the tube: inside the wave pays, but the foam drags you back; pump to get out before it carries you to the edge. In the air, circle the mouse around conti (drag a finger, or ← →) to flip, and land on your feet: upside down costs you a board. Skimboarding: a flat board with no fins.
Notes fall down the lanes: five on a computer (A S J K L or 1 to 5), three wider ones on a phone (just tap). Hit them inside the glowing zone on the line. Long notes have a tail: keep holding until it ends (in Vai mode, the whammy bends them). Hits fill the Vai meter. When it’s full, press Space, the Vai button or the whammy bar: Vai mode adds overdrive and doubles your points. Steve Vai is Leo’s all-time idol. He wanted his favorite Vai song here, but copyright got in the way, so the game plays an original tribute. The real one is legendary: For the Love of God ↗
Monday to Thursday, 7 to 9 p.m., Leo is on the mats. Your turn: roll with a bug.
Pick the move, then hit Now! on time (Space, Enter or tap). Points like on the mats: guard pass 3, mount 4, sweep 2, submission wins.
Bash(open mailto:)
contact
If you build agents too, let’s compare notes.
Leo is always up for a conversation about agents, data and what actually works in production. Email or LinkedIn is the fastest way.
hi. this is conti, Leo’s agent.
type a command, ask me something, or just use the buttons.