Sergio Francisco

Kojo ran its services on EKS with logging that was hard to query and no distributed tracing. I led the rollout of OpenTelemetry as the platform standard.

I designed a hybrid collector architecture, with an agent running as a DaemonSet on every node feeding a gateway collector running as a Deployment, then presented it to the team, collected feedback, and secured approval before building anything.

From there, I ran the project end-to-end: validated auto-instrumentation in staging, configured the collection, processing, and export pipeline to Datadog, verified that resource tagging was applied correctly, and confirmed the Datadog API integration. Then, reproduced the same approach in the production environment.

The outcome was distributed tracing across staging and production with near-zero application code changes, so service teams got traces without touching their own code.

I self-managed the project in Jira on an Agile cadence, defining my own sprints and publishing a weekly plan, progress, and blockers to the wider team.

CoinList modernized its infrastructure by migrating six Ruby on Rails applications from EC2 to AWS ECS Fargate, using Docker for containerization, Terraform for IaC, and GitHub Actions for CI/CD. This significantly reduced costs, improved scalability, and accelerated deployment processes. The new platform is more resilient and efficient, allowing CoinList to handle traffic spikes during token sales and streamline operations.

Check all details of this project by clicking on the case study link.

Sight Machine offered NVIDIA Omniverse as a digital twin capability alongside its manufacturing analytics platform, deployed into each customer's own Azure account. I took over the project after the engineer running it left, inheriting Terraform modules based on Microsoft's Omniverse Azure Operations Twin reference.

I received it with one customer in production and left it with roughly ten enterprise deployments across the automotive and beverage sectors. I evolved the inherited modules, fixing defects and reducing dependencies: they provisioned the full Azure footprint (VNet, AKS, APIM) and the Omniverse workloads themselves, with Flux installed and driven from the same Terraform, from a per-customer monorepo with remote state in Azure Blob Storage.

The project had no GPU observability, so I added it. NVIDIA's dcgm-exporter fed the GPU nodes into in-cluster Prometheus, driving Grafana dashboards and AlertManager alerts on thresholds I wrote for utilization, memory, errors, and temperature, routed to Slack and escalated to PagerDuty. With deployments in customer accounts rather than ours, that telemetry made remote-assisted operations possible.

Before rolling off, I trained the two engineers who took it over.

Kojo's main monolith ran CI on CircleCI, and the pipeline duration was slowing developer throughput. I migrated it to GitHub Actions across non-trivial release workflows.

Instead of throwing larger machines at the problem, I moved jobs to memory-optimized Graviton instances (r8g.4xlarge) and tuned Node.js heap allocation per step. Pipeline duration dropped roughly 50% at near-neutral cost, around $2 per full CI run: a more expensive instance running for far less time. I defended the instance choice in the developer experience guild once the numbers confirmed it.

The CircleCI license was retired, and developer experience scores improved.

I converted the entire workflow and drove most of the test iterations using Claude as an AI-assisted development tool, which brought the migration in on deadline.

Web3 Pro is a B2B SaaS platform in the Web3 space. Its AWS estate had grown beyond what a single account could reasonably govern, and it needed an account strategy that would withstand security and compliance reviews.

I ran a security assessment of the existing environment, designed the AWS Landing Zone, and implemented AWS Control Tower with an Account Factory for Terraform, enabling new accounts to be created and customized from code within the organization's security guardrails rather than by hand. I enrolled the existing accounts into the new structure and delivered the training and documentation the team needed to operate it without me.

The result is a multi-account foundation where account creation is repeatable, auditable, and governed by default.

An internal Toptal product: a platform that remediates documents to WCAG 2.2 AA accessibility compliance, now running in production.

I was the sole infrastructure engineer on a three-person team. I designed and built the full AWS footprint as reusable Terraform modules across staging and production, covering multi-service ECS Fargate for the API, front end, workers, and SSO, as well as RDS, S3, SQS, SES, and ACM.

The platform runs on a private VPC with interface endpoints for ECR, Secrets Manager, CloudWatch Logs, ECS, and SQS, keeping service traffic off the public internet.

Delivery runs on GitHub Actions with AWS OIDC federation, which removed long-lived cloud credentials from the deployment path entirely.

EVgo operates a public EV charging network in the United States. They wanted three applications consolidated from Rackspace to AWS with minimal downtime.

I ran the discovery to map dependencies, then chose re-platforming over a straight lift-and-shift. I improved the Dockerfiles, built the CI/CD pipelines with Bitbucket Pipelines, configured Kubernetes objects, and deployed workloads to Amazon EKS, with the static front end served from S3 via CloudFront.

Traffic was migrated during a scheduled maintenance window. The applications ended up in a single consolidated AWS environment with automated delivery, replacing the manual deploys they had on the previous provider.

TeleTracking is a healthcare software company. Its Azure footprint had become expensive and difficult to operate, and the company wanted to migrate those workloads to AWS.

I assessed the existing environment, conducted a total cost of ownership analysis to support the migration decision, and designed the target network architecture: a hub-and-spoke topology spanning three regions across the United States and Europe, centralizing ingress and egress through AWS Transit Gateway, with WAF and load balancers at the edge.

My part of the project was delivered in one month.

The client, a Brazilian fintech company, wanted to migrate its infrastructure from an on-premises data center (TIVIT) to a public cloud provider (AWS) using the re-platforming technique.

As the architect, I first evaluated the technical and financial aspects of several cloud vendors and ultimately chose AWS as the platform. Next, I reviewed their entire legacy infrastructure, designed a multi-account/region/VPC architecture, and collaborated with the engineering team to migrate a set of systems that processed millions of financial transactions daily.

This migration brought order to the client's infrastructure architecture and operations, which had been in a state of chaos. As a result of this project, I received an invitation from AWS and Claranet, an AWS partner, to present the migration case at AWS Summit São Paulo 2017.

4 Elements Music, a music platform, hired me to revamp their infrastructure using AWS ECS and Fargate to containerize a Python/Django application on AWS.

The old system lacked scalability and performance. The solution involved containerizing the app with Docker, building a new platform with Terraform, deploying secure networking with VPC, and automating software delivery with CI/CD.

This boosted efficiency and security and laid a solid foundation for the new platform's launch.

Whatnot's recommendation engine was trained and served on SpellML. The company wanted it on Amazon SageMaker to mature its MLOps practice.

I designed the SageMaker infrastructure, built the training pipelines, deployed the models as real-time endpoints, and integrated them with the client's existing MLOps tooling.

Moving off SpellML put both training and serving on managed AWS services, so the team could iterate on models without operating the underlying platform.

Art of Problem Solving is a math education company whose application stack predated its growth. Caylent was engaged to show how that stack could be modernized on AWS.

I delivered a proof-of-concept for a container-based platform on ECS Fargate, with CircleCI and GitHub handling build and delivery, so the team could evaluate the target architecture against their existing stack before committing to a full migration.

The proof of concept covered container orchestration, the delivery pipeline, and the operational model, giving AoPS a concrete basis to decide on modernization rather than an estimate.