Skip to content
Renan Sued

Rio de Janeiro, Brazil · UTC−3 · Remote or relocation

Renan Sued

Senior AI Engineer · Solution Architect

I build production software: enterprise platforms, AI agents, and the engineering teams that run them.

I've shipped software since 2014, from mobile apps to Java and AWS enterprise platforms. Today I lead teams whose platforms reach 93.4M+ six-month active users. Since 2020 I've built LLM agents with guardrails, retrieval and human review, at work and in BikerWay, my own product.

Explore my workDownload CV
~/renan

whoami

Renan Sued

Senior AI Engineer · Solution Architect

Rio de Janeiro, Brazil · UTC−3

open to remote or relocation

Engineering impact

Scale I've worked at, not adjectives.

Every number here comes from my resume or my own records. Where there is no number, there is no claim.

  • shipping software since

    2014

    About 12 years, from mobile apps to enterprise platforms.

  • six-month active users

    93.4M+

    Users active within a six-month window on the enterprise platforms delivered by the teams I lead.

  • teams · engineers led

    5 · 15

    Cross-functional teams at Lumis, plus mentoring engineers and architects.

  • of daily transactions

    Millions

    Through the APIs, event-driven services and DXPs I architected.

  • AI systems delivered at Lumis

    4

    SDLC agents, process-automation agents, a RAG assistant, a content agent.

  • hands-on with LLMs since

    2020

    Claude, OpenAI, Gemini and DeepSeek models.

What I build

Six kinds of systems, one way of working.

Most of my career has been enterprise backend and platform work. AI and product engineering are where that foundation is going.

  • AI Systems

    Agents and LLM applications built for real workloads: SDLC automation, process-automation agents with prompt-injection protection and a retrieval-backed assistant (both in production), and content agents with human review.

    GuardrailsRAGHuman-in-the-loopMCP

  • Distributed Systems

    Event-driven services and high-availability platforms for regulated financial clients, and event queuing with guaranteed execution and no duplicate processing.

    Event-drivenHigh availabilityKafkaPython

  • Backend Platforms

    Java APIs and Digital Experience Platforms taken from design to deployment, processing millions of daily transactions with security and compliance built in.

    JavaSpring BootREST APIsDXP

  • Cloud Infrastructure

    Cloud-native architecture on AWS with infrastructure as code, containers and least-privilege access. Managed Postgres, edge functions and serverless when they fit the constraint better.

    AWSIaCDockerSupabase

  • Developer Platforms

    The machinery that lets teams ship: DevSecOps and LLMOps, CI/CD, quality gates, OpenTelemetry, and agents that take a task from pickup to staging.

    CI/CDSonarQubeOpenTelemetryClaude Code

  • Product Engineering

    End-to-end products: BikerWay across mobile, web, admin and data; A/B and multivariate experimentation for marketing platforms; fintech products at Sued Team, which I founded.

    MobileWebExperimentation0 → 1

AI in production

The model is one box. The system is the job.

Every AI system I ship follows the same path from intent to production. Select a step to see what it means in practice and where I've built it, with an honest status label on each piece.
Production
Runs against real users or production data
In use
Internal tooling I actively use
Experimental
Built and tested, not operated regularly yet
Planned
Designed and on the board, not built

A six-step pipeline. Select a step to see what it means and where I have built it.

  1. 01 · intent & approval

    Human

    Sets the goal and owns the risk. Anything irreversible, paid or public waits for an explicit yes.

    Where I've done this

    • Content agent with human-in-the-loop review

      Production

      Confidence scoring decides which items a person reviews before publishing.

    • Approval gate before any paid pipeline run

      Production

      BikerWay data pipelines stop and ask on Telegram before paid or slow work.

    • Human-only risk gates for the agent team

      In use

      Merge to main, production migrations, first paid API call and public publishing always stop for me.

The concepts, grounded

  • Agent orchestration

    One coordinator, many narrow specialists, explicit rules for who decides when they disagree.

    Lumis AI-first delivery · BikerWay agent team

  • Tool calling & MCP

    Agents act through tools with scoped credentials, not through free-form access.

    Lumis tooling · BikerWay pipelines

  • Structured outputs

    Model output is data with a schema. If it doesn't parse, it doesn't ship.

    BikerWay data platform

  • RAG

    Chunking, embeddings and vector search behind an assistant used by a law firm's clients.

    Lumis · multinational law firm

  • Automated workflows

    Agents that move a task from pickup to staging, or content from search to publish.

    Lumis SDLC agents · content agent

  • Human-in-the-loop

    Confidence decides what a person must review; humans approve anything paid, public or irreversible.

    Lumis content agent · BikerWay approval gates

  • Observability

    Traces, run histories and cost per call, so an agent's behaviour can be explained after the fact.

    OpenTelemetry at Lumis · BikerWay pipeline runs

  • Evaluation & quality gates

    Scores computed from objective signals, with thresholds per domain that decide auto-approve or review.

    BikerWay ingestion layer · Lumis content agent

  • Guardrails

    Input/output validation and prompt-injection protection. Third-party content is data, never instructions.

    Lumis production agents

  • Production reliability

    CI/CD, least privilege, budgets that fail closed and approval gates in front of paid or risky steps.

    Lumis LLMOps · BikerWay runtime

Case study · BikerWay

My own product, built like a production system.

A map, convoy and club platform for motorcyclists, with AI pipelines that keep its catalogue of routes and places alive.
BetaAndroid beta · iOS laterFounder & lead engineer, with a small team

The problem

Riders in Brazil plan trips around questions generic map apps don't answer well: which fuel stations, workshops and places to stay can be trusted, and which roads are worth riding on a motorcycle. Clubs and group rides run on scattered chat groups. No single product brought the map and the community together.

The goal is something like PlugShare for motorcyclists: a shared map that the community confirms, corrects and extends, where a new rider sees useful routes and places on the first day.

The product

  • Live convoys

    Group rides on a live map. A native background-location task keeps the convoy visible while phones sit in pockets.

  • Routes & trusted places

    Recommended routes, rider-relevant places (fuel, workshops, viewpoints), reviews and a community consensus on fuel prices.

  • Clubs

    Motorcycle clubs with rankings, territories, recruitment and check-ins.

  • Events & invites

    Events and club meetups, shareable invite links that open the app or fall back to a download page.

  • Garage

    A virtual garage with fuel consumption, expenses and maintenance history per motorcycle.

  • Ride journal

    A log of past rides with distance, duration, number of motorcycles and photos.

AI inside it, with honest status

  • Route & place discovery pipelinesProduction
  • Single ingestion layerProduction
  • Agent development teamIn use
  • AI marketing engineExperimental
  • AI runtime with cost controlExperimental
  • Content agent team (AG-0…AG-7)Planned
Read the full case studybikerway.com.br
BikerWay road view: a dark map around Rio de Janeiro with colored route lines, place pins and a card for the Rio to Petrópolis challenge route.
Create-convoy form: starting point chosen from an event, a destination place, departure date and time, and privacy options open, private or club-only.

AI workforce · my working thesis

Agents as an engineering organization, with a human holding the keys.

Generating text is the easy part. The interesting engineering is designing responsibilities, tools, permissions, memory, workflows, evaluation, human approval, observability and failure handling. I test that thesis on real work with a team of 14 Claude Code agents that refines and builds BikerWay's roadmap.

The BikerWay agent team, from the human founder down to specialists. Select an agent to see its role.

Human

Orchestration

tech-lead

In use

Orchestration

Reads the story, picks the relevant specialists, merges their positions into one plan and keeps architecture docs and the board honest.

Synthesizes when positions converge.

Arbitration

Product

Build

Quality & risk

What this is, honestly

This is not an autonomous company. The agents don't talk to each other in the background and nothing runs on a schedule yet. A session I open invokes each specialist, collects their written positions and applies the arbitration rules. The value is in the constraints: who may decide what, what needs a human, and what gets written down.

How a story becomes code

  1. 1

    Story. I write or update the story on the board.

  2. 2

    Triage. The tech lead decides which specialists matter for it.

  3. 3

    Consult. Each specialist returns a short written position: approach, risks, cost.

  4. 4

    Converge or arbitrate. Agreement → one plan. Technical split → CTO. Business split → CEO. Vulnerability → security veto.

  5. 5

    Implement. Build agents execute the single plan.

  6. 6

    Verify. QA runs the suites and lists what still needs a real device.

  7. 7

    Review. Security (blocking) and legal review anything touching auth, personal data or payments.

  8. 8

    Close. The board is updated to match reality, not intentions.

Risk gates that always stop for a human

  • Merge to main / production deploy
  • Applying a schema or RLS migration in production
  • The first paid API call in a session, and any large batch
  • Publishing anything publicly
  • Using a personal token or rotating credentials
  • Switching LLM provider in the middle of a story

The design problem, dimension by dimension

Production
Runs against real users or production data
In use
Internal tooling I actively use
Experimental
Built and tested, not operated regularly yet

  • Responsibilities

    In use

    One role per agent, written down as a versioned definition, with explicit decision rights.

  • Tools

    In use

    Each role's definition scopes it to its own repositories. Next: content agents that act only through an MCP server over one ingestion layer.

  • Permissions

    Production

    The database checks permissions, not the prompt. A scoped service account per pipeline today; one per content agent is planned.

  • Memory

    In use

    A shared context repo (architecture, schema checklist, roadmap, real status log) is loaded into every session in every repo.

  • Workflows

    In use

    The consensus protocol for building; discover → enrich → score → ingest for data.

  • Evaluation

    Production

    Confidence scored from evidence. Next: per-agent approval rate and cost per approved item in the curation queue.

  • Human approval

    In use

    Risk gates for anything irreversible, paid or public; an approval message before every paid pipeline run.

  • Observability

    Production

    Run and step history for every pipeline; token cost recorded per call.

  • Failure handling

    Experimental

    Idempotent resume by status, budgets that fail closed, and a planned per-agent off switch.

Next layer: content agents in the production data path

Planned

A team of agents that keeps BikerWay's catalogue alive. Each will have its own database identity and permissions, write only through the shared ingestion layer, and send anything below its confidence threshold to a human curation queue.

  • AG-0 Editor-in-chief
  • AG-1 Places
  • AG-2 Events
  • AG-3 Famous routes
  • AG-3B Composed routes
  • AG-4 Challenges & badges
  • AG-5 Parts & manuals
  • AG-6 Offers
  • AG-7 Verifier

Experience

From shipping apps to leading platform teams.

Twelve years, two arcs: growing technical responsibility at Lumis, and AI moving from tooling to the systems I build.

  1. Dec 2021 – present · Rio de Janeiro

    Solution Architect | Software AI-Systems Engineer · Lumis

    Technical lead for five teams building enterprise platforms, and the engineer behind the AI systems around them.

    Impact

    • Lead 5 engineering teams (15 engineers) delivering enterprise platforms for 93.4M+ six-month active users; mentor engineers and architects; chapter leader for practices and tooling.
    • Architected scalable APIs, event-driven services and multiple Digital Experience Platforms from design to deployment, processing millions of daily transactions.
    • Led engineering and architecture for critical banking and insurance clients: security, compliance and high availability for regulated workloads.
    • Designed developer agents that automate the team's SDLC end to end, from task pickup through implementation and testing to staging handoff.
    • Built and operate production AI agents for process analysis and solutioning, with LLM guardrails, input/output validation and prompt-injection protection.
    • Led a WordPress → Java/Lumis XP migration for a multinational law firm, including a production AI assistant built on vector databases, embeddings and chunking.
    • Delivered an autonomous content agent (search, generate, analyze, publish) with confidence scoring and human-in-the-loop review.
    • Established DevSecOps and LLMOps practice: CI/CD, infrastructure as code, Docker, SonarQube, OpenTelemetry and least-privilege access.

    Key engineering challenges

    • Making AI trustworthy enough for production: validation, guardrails and human review designed in from the start.
    • Security, compliance and availability requirements of regulated financial clients.
    • Bringing AI-first tooling to five teams without fragmenting how they work.

    Technologies

    JavaAWSCloud-nativeAgent orchestrationMCPVector databasesDockerIaCSonarQubeOpenTelemetryClaude CodeCursorLumis XP

  2. Jan 2017 – Dec 2021 · Rio de Janeiro

    Software Development Engineer · Lumis

    Full-stack engineer who grew into technical leadership of a major retail web platform and of the team's DevOps modernization.

    Impact

    • Technical leader for the web platform of a major Brazilian shopping-mall group: corporate applications on content experimentation platforms with Java, React, Vue.js, AWS and SQL.
    • Technical leader for DevOps modernization (Git, Jenkins, Docker, SonarQube) and event-driven architecture; automated A/B, auto-targeting and multivariate testing for marketing experiments.
    • Event queuing and file processing with Python and Kafka, with guaranteed execution and no duplicate processing.
    • Full-stack development across Java, JavaScript, Node, React, Redux, Lumis XP, jQuery and XSL.

    Key engineering challenges

    • Introducing CI, containers and code-quality gates into a delivery process that was already running.
    • Automating marketing experiments: A/B, auto-targeting and multivariate tests.

    Technologies

    JavaJavaScriptReactVue.jsNodePythonKafkaAWSSQLJenkinsDockerSonarQube

  3. Jun 2015 – Dec 2016 · Rio de Janeiro

    Software Engineer · Founder · Sued Team

    Founded the team and led engineering for enterprise fintech applications.

    Impact

    • Built end-to-end fintech products with Java, Android, MVC, MySQL, Node and React.
    • Managed 4 projects across military, pet-tech and political-research domains.

    Key engineering challenges

    • Owning a product from first line of code to client delivery.

    Technologies

    JavaAndroidMySQLNodeReact

  4. Jan 2015 – Dec 2016 · Rio de Janeiro

    Software Engineer Mobile · Fluxo Consultoria UFRJ

    Mobile apps with security and recognition requirements.

    Impact

    • Military fitness-assessment app for FAB (Brazilian Air Force): Android, Firebase, JWT, OAuth2 and custom cryptography.
    • Automated medicine recognition from prescriptions with OCR (Tesseract).

    Key engineering challenges

    • Security requirements of a military client on a mobile platform.

    Technologies

    AndroidFirebaseJWTOAuth2Tesseract OCR

  5. Jun 2014 – Dec 2016 · Rio de Janeiro

    Software Engineer Fullstack · Sued Team

    Where it started: apps for students and teachers.

    Impact

    • Mobile and web apps for students and teachers at two federal public education institutions (MySQL, PHP, GCM push, Android, Linux VPS).

    Key engineering challenges

    • Shipping and hosting real apps for real schools while still studying.

    Technologies

    AndroidPHPMySQLGCMLinux

Education & credentials

  • B.Eng. Control & Automation Engineering, PUC-Rio · 2024

    Additional domain in Software Engineering. Thesis: GADEMO Web, simulation and evolutionary optimization with genetic algorithms.

  • Technical Diploma, Control, Robots & Automation, CEFET/RJ

    Secondary-level technical program.

  • Google Generative AI Leader · Google Cloud · 2026

    Certification in strategic and applied generative AI.

  • Claude Certified Architect · Anthropic · in progress

    Not obtained yet.

  • 2nd place, PUC-Rio Sustainability & Energy Hackathon · 2018

Selected engineering work

Problems, decisions and results.

Client names stay confidential; the engineering doesn't. Fields with nothing verified behind them are left out rather than filled in.

  • My product · founder & lead engineer

    BikerWay

    Beta

    Riders lack a trustworthy, rider-specific map of routes and places, and the tools for group rides and clubs are scattered.

    Approach, architecture, result & lessons
    Approach
    A native app on a shared Supabase backend, a data platform where AI proposes catalogue entries under rules and human approval, and a team of Claude Code agents that refines and builds the roadmap.
    Architecture
    Five repositories, one Supabase project, one ingestion layer, approval gates in front of paid and irreversible work.
    Technology
    Expo / React NativeTypeScriptSupabaseNext.jsViteZodn8nGemini · OpenAI · DeepSeekClaude Code
    Result
    Android app in beta with real testers; web app and admin panel in production; AI pipelines writing evidence-backed data to production.
    What I learned
    The hardest cross-repo bugs came from missing shared context, not from having many repos. Context is infrastructure.
    Read the case study

  • Lumis · 2021 – present

    AI systems at Lumis

    Delivery teams and enterprise clients needed AI that holds up in production, with validation, guardrails and human review.

    • SDLC developer agents

      In use
      Problem
      Manual hand-offs between picking up a task and handing it to staging.
      Approach
      Autonomous developer agents covering implementation and tests up to the staging/homologation handoff.
      Result
      Manual steps removed from the release path; releases accelerated.

    • Process-automation agents

      Production
      Problem
      Process analysis and solutioning done by hand.
      Approach
      Production agents with agent training, LLM guardrails, input/output validation and prompt-injection protection.
      Result
      In production, built and operated by me.

    • Assistant for a multinational law firm

      Production
      Problem
      A WordPress platform that had to scale, and a client-facing assistant on top of it.
      Approach
      Migration to Java / Lumis XP at scale; retrieval with vector databases, embeddings and chunking.
      Result
      Production AI assistant for client-facing use cases.

    • Autonomous content agent

      Problem
      Registering content on emerging platforms.
      Approach
      Search → generate → analyze → publish, with confidence scoring deciding when a human reviews.
      Result
      Delivered with human-in-the-loop review.
    Technology, result & lessons
    Technology
    AI agentsAgent orchestrationMCPVector databasesEmbeddingsLLM guardrailsJavaAWS
    Result
    Four AI systems delivered; two of them run in production.
    What I learned
    In production AI, the validation and review paths take more design work than the prompt.

  • Lumis · banking and insurance clients

    Regulated platforms & DXPs

    Production

    Critical banking and insurance clients need platforms that are secure, compliant and always available.

    Approach, architecture, result & lessons
    Approach
    Scalable APIs, event-driven services and Digital Experience Platforms, designed and taken to deployment, with DevSecOps practices and least-privilege access.
    Architecture
    Cloud-native services on AWS; event-driven integration; CI/CD with quality and security gates.
    Technology
    JavaAWSEvent-driven architectureDockerSonarQubeIaCOpenTelemetry
    Result
    Millions of daily transactions with high availability, performance and security for regulated workloads.
    What I learned
    In regulated systems, compliance is an architecture input, not a review step at the end.

  • Open source · Java

    Distributed rate limiter

    Enforce a global rate limit across a fleet of servers without a network call on every request.

    Approach, architecture, result & lessons
    Approach
    Sharded counters per key on a distributed key-value store, local batching flushed on an interval, and non-blocking decisions from the last known global count plus pending local deltas.
    Architecture
    Store abstraction, retry policy and circuit breaker for store failures, pluggable metric publishers (Prometheus, CloudWatch).
    Technology
    JavaCompletableFuturePrometheusAWS CloudWatchMaven
    Result
    Public repository with a builder API, a mock store for tests and documented operational trade-offs.
    What I learned
    Choosing eventual consistency for latency is a product decision. The error bound should be explicit and observable.
    View on GitHub(opens in a new tab)

  • FAB (Brazilian Air Force) · via Fluxo Consultoria UFRJ

    Military fitness-assessment app

    A mobile fitness-assessment app for the Brazilian Air Force.

    Approach, architecture, result & lessons
    Approach
    Native Android app with Firebase, token-based auth (JWT, OAuth2) and custom cryptography.
    Technology
    AndroidFirebaseJWTOAuth2Cryptography
    Result
    Delivered mobile app for the Brazilian Air Force.
    What I learned
    Security requirements shape the data model from day one, not the release checklist.

  • B.Eng. thesis · PUC-Rio · 2024

    GADEMO Web

    Simulation and evolutionary optimization with genetic algorithms, one of the foundations of machine learning.

    Approach, architecture, result & lessons
    Approach
    A computational approach to simulating genetic algorithms and evolutionary optimization, presented as my engineering thesis.
    Technology
    Genetic algorithmsEvolutionary optimizationSimulation
    Result
    Bachelor of Engineering thesis, PUC-Rio.
    What I learned
    Optimization is search under constraints, the same framing I now use for agent workflows.

Engineering stack

Depth over logos.

Grouped by what the tools are for. Solid chips are tied to a role or project described on this page; dashed ones I've used, without a project to show here.
Shipped tied to work on this pageUsed experience without a published project

  • Languages

    Java for a decade of enterprise work, TypeScript for everything BikerWay.

    • Java
    • TypeScript
    • Python
    • JavaScript
    • SQL
    • Kotlin
    • Go

  • Backend

    APIs and platforms built to sit behind products other people depend on.

    • REST APIs
    • Node.js
    • Lumis XP (DXP)
    • Edge Functions
    • Spring Boot
    • Hibernate / JPA

  • AI

    Agents and LLM apps with the parts that make them safe to run.

    • AI agents
    • Agent orchestration
    • MCP
    • RAG · embeddings · chunking
    • Guardrails
    • Prompt-injection defense
    • Human-in-the-loop
    • Structured outputs (Zod)
    • LLMOps
    • Claude Code · Cursor
    • Claude · OpenAI · Gemini · DeepSeek
    • pgvector · Pinecone · Chroma
    • LangChain

  • Cloud

    AWS for enterprise; managed Postgres and edge platforms when the budget is the constraint.

    • AWS
    • Supabase
    • Vercel
    • Firebase
    • Serverless / Lambda
    • Azure
    • Cloudflare

  • Infrastructure

    Pipelines that make the safe path the default path.

    • Docker
    • Infrastructure as Code
    • CI/CD
    • Jenkins
    • SonarQube
    • DevSecOps
    • Kubernetes
    • Terraform
    • GitLab CI

  • Data

    Events, relational data and the provenance that keeps them trustworthy.

    • Kafka
    • Postgres + RLS
    • MySQL
    • Vector databases
    • OpenStreetMap · Google Places
    • Elasticsearch
    • NoSQL

  • Observability

    If an agent did something, I want to be able to explain why.

    • OpenTelemetry
    • PostHog
    • Sentry
    • Pipeline run tracing
    • Prometheus
    • Grafana
    • Datadog

  • Architecture

    Patterns chosen for the constraint, not for the résumé.

    • Event-driven architecture
    • Cloud-native
    • High availability
    • Digital Experience Platforms
    • Experimentation (A/B, multivariate)
    • Microservices
    • Distributed systems

How I think

Principles I build by.

Not slogans. Each one is tied to a system where I apply it, named below it.
  1. 01

    Build systems, not demos.

    A demo answers "can it work?". A system also answers "what happens when it doesn't?". Guardrails, validation, observability and rollback come with the feature, not after it.

    from: Production agents at Lumis

  2. 02

    AI is an engineering problem, not only a prompt.

    Schemas, retrieval, evaluation, permissions and cost all matter more than the wording of a prompt once real users depend on the output.

    from: RAG assistant · BikerWay data platform

  3. 03

    AI proposes, rules decide, humans curate.

    The model suggests. A deterministic score built from evidence decides what is safe to publish, and a person reviews the rest. The model never approves its own work.

    from: BikerWay agent design

  4. 04

    Third-party content is data, not instructions.

    Anything an agent reads from the outside world may be hostile. Prompt-injection protection belongs in the architecture, enforced where the prompt can't override it.

    from: Guardrails at Lumis · BikerWay ingestion

  5. 05

    Architecture exists to solve constraints.

    Regulated finance made availability and compliance the inputs. A fixed monthly budget made free tiers and fail-closed AI spending the inputs. Same discipline, different constraints.

    from: Banking & insurance platforms · BikerWay

  6. 06

    Automation needs provenance and an off switch.

    Every automated write should say who wrote it, from what evidence, and be stoppable without a deploy. That is what makes automation something a team can actually trust.

    from: BikerWay ingestion layer

Off the clock

I like building things that move.

My engineering training is in control, robotics and automation: systems that sense and act. I ride, and BikerWay is what I build for riders. Trains, travel and electronics keep me curious about how physical systems are designed and run.

  • Motorcycles

    Riding, and building BikerWay for other riders.

  • Railways

    Large, scheduled, safety-critical systems moving in real time.

  • Electronics & IoT

    Sensors, microcontrollers, and the line between hardware and code.

  • Travel

    New places, new constraints, and a reason to plan routes.

Contact

Have a hard engineering problem?Let's talk.

Rio de Janeiro, Brazil · UTC−3 · Remote or relocation. Open to AI Engineer, Applied AI, Forward Deployed, AI Platform and Staff / Senior Backend roles.

renansued3@gmail.comLinkedInGitHubCV (PDF)