Miguel Pereira

Miguel
Pereira

I lead a cloud, data and AI practice, and I still keep my hands on the keyboard.

I started out in 1998 as a sales engineer, moved to the US in 2001 and worked my way from sysadmin to DBA to DevOps to building data and AI platforms. These days I lead a consulting practice that helps organizations get their data in order and put AI to work in production. I still architect, review and build, and I'm a perennial tinkerer of all things new and shiny: networking, data platforms, applied AI, process optimization and whatever idea I can't stop thinking about.

Practice Director, Cloud Acceleration at Protagona · 7 AWS certifications earned since 2016

Modernizing a data platform, trying to get GenAI out of the demo stage, or putting AWS to work for a nonprofit? Let's talk. Message me on LinkedIn

2001Oblio2011ExamSoft2018Onica2024Director

What I doFour layers, from the ground up

The way I see it, AI is only as good as its data, the data is only as good as the platform and cloud underneath it, and none of it matters if you can't trust it.

Surface: AI in production

  1. Layer 1

    Cloud and network foundations

    The migrations, networking, Kubernetes and disaster recovery that keep data and AI workloads reliable without blowing the budget, from multi-account AWS topology all the way out to the edge.

    EKS and Kubernetes, Terraform and CDK, Transit Gateway and multi-account networking, Well-Architected

  2. Layer 2

    Security by default

    I'd rather build security in than bolt it on later: identity and least privilege, edge protection, secrets and the supply chain, and defenses for LLM applications.

    IAM and cross-account trust, Cognito and OAuth, WAF and CSP, KMS and Secrets Manager, SBOM and vulnerability scanning, Prowler

  3. Layer 3

    Modern data platforms

    Serverless lakehouses on AWS, end to end: ingestion, Iceberg medallion layers, governance, quality checks and serving. Plus the unglamorous part, getting the data off legacy databases and out of data centers to land there.

    S3 and Iceberg, Glue and Athena, DuckDB, DMS with CDC, Databricks, RDS and Aurora

  4. Layer 4

    Applied AI

    Getting GenAI past the demo and into production: agents, evaluation, retrieval and document intelligence. I keep a close eye on cost, accuracy and safety, and reach for a fine-tuned small model when a big one is overkill.

    Bedrock, AgentCore and Strands, MCP, RAG and GraphRAG, LLM-as-judge evaluation, LoRA fine-tuning, OCR and vision-language extraction, agentic Amazon Connect

Bedrock: the cloud it all runs on ↓

Now2024–present

Practice Director, Cloud Acceleration at Protagona

A practice is only as good as its people, and I get to lead 25+ incredibly talented architects and engineers. My job is to help them grow, put them on work that stretches them and back them up with the offerings, partners and tools they need. I haven't let go of the keyboard either: 40 client engagements since 2021.

  • OfferingsBuilt the practice's repeatable offerings and still run them: AWS-funded GenAI assessments, plus accelerators for data lakes, AI document processing and Amazon Connect contact centers.
  • CredentialsLead the work that earns and keeps Protagona's AWS competencies.
  • EnablementEnable the team to grow across AWS, Google Cloud and Databricks through hands-on cohorts.
  • OperationsHelp run the company's operations alongside the other delivery leads: staffing, delivery health, and the data and tools leadership uses to steer the business.
  • Deal to deliveryTurn what client executives actually want into scoped, AWS-funded statements of work, and set the qualification rules that keep risky deals out.
  • FlagshipLead delivery of a youth career-coaching AI platform, from the executive relationship and architecture direction down to hands-on guardrail and security work.
  • AI-native deliveryBuild the platforms the practice runs on: the lakehouse behind our KPIs, plus tools for staffing, estimating, assessments, funding, competencies and access audits. Co-built the presales and account-health platforms too.

SegmentsA few highlights, client names withheld

Applied AI

  • Agents in prod2026

    Shipped production agents on Bedrock AgentCore and Strands, both scheduled and tool-calling, and every one had to pass its evaluations before release.

  • Rubric, not gut2026

    Picked the production model for a youth career-coaching AI by scoring the candidates against a rubric instead of a hunch. The winner more than doubled the runner-up on grounding at nearly the same latency.

  • Agentic CX2026

    Built an agentic contact-center accelerator on Amazon Connect: AI self-service, escalation with agent assist, Contact Lens and budget controls, all deployed from a single configuration and proven in a live environment.

  • $0.0392026

    Per minute of footage: what GenAI transcription cost a nonprofit film archive, at 93–98% accuracy.

  • Idea → roadmap2025

    Through funded assessments, walked nonprofits, public media, health, hospitality and SaaS organizations from a GenAI idea through model evaluation to a production roadmap.

  • Price vs. accuracy2025

    Built the evaluation harness behind those assessments. It scored Bedrock and local models on public and finance benchmarks, so clients could weigh price against accuracy before committing.

The climbWhere the data instinct comes from

2021–24Protagona: Practice ManagerAWS Advanced Tier partner in Plano, Texas: data engineering, AI, cloud acceleration and DevOps.
  • Architect on a multi-year AWS operations and modernization program for a US digital bank: reduced its security-findings backlog, automated how patches were applied across production and non-production and moved its crypto platform's database to Amazon RDS
  • Scoped Migration Acceleration Program assessments and a network-firewall modernization for a data and analytics firm
  • Delivered AWS networking immersion training, ran the practice's EKS launch offering (standing up clusters and migrating workloads) and led Well-Architected reviews through to remediation
2018–21Onica / Rackspace: Cloud Engineer to Practice ManagerOnica was a large AWS Premier Consulting Partner and managed services provider, acquired by Rackspace Technology in 2019.
  • Built a production GemFire key-value data store on AWS
  • Led automation and instrumentation for a major US airline's migration of its distributed monolith to EKS
  • Migrated a media company's Windows/.NET workloads to AWS
2011–18ExamSoft: Senior DBA to Senior DevOps EngineerSecure exam software used by law, nursing and medical schools; acquired by Turnitin in 2020.
  • Ran 1,300+ production databases across MSSQL failover clusters
  • Moved on-prem products to AWS with CloudFormation, Ansible and Packer
1998–2011Early career: Oblio Telecom and VerizonOblio was a Dallas distributor of international prepaid calling cards that launched its own prepaid mobile service in 2005.
  • Started in 1998 as a sales engineer and moved to the US in 2001
  • At Oblio: grew from sysadmin into launching its direct-to-consumer e-commerce platform, running its contact-center infrastructure and building back-office data and reporting
  • Architected the MVNO program integrating with Sprint and Telcordia
  • At Verizon: consulting DBA and systems architect

7 AWS certifications earned between 2016 and 2026, including Solutions Architect Professional and DevOps Engineer Professional.

Off the clockEvery hobby eventually gets a repo

Guitars4
  • guitarserials.orgFree guitar serial decoder: 18 makers, 115 format rules, each traced to at least two sources.
  • Live guitar detectorPoint a phone camera at a guitar and it tells you the brand and model in about two seconds, no GPU needed.
  • Guitar CollectionA guitar catalog that uses Bedrock to pull specs out of PDFs and maker sites.
  • StemForgeSelf-hosted stem separation with a multitrack mixer, chord detection and synced lyrics.
Cycling3
  • Training AppLocal-first indoor trainer app that speaks BLE, plus ANT+ straight over USB. Rust, Tauri, Svelte.
  • Bike geometryMeasures frame geometry from a photo, using the wheel's standard rim size as the ruler.
  • Bike FinderWhy hunt for bikes by hand when Python and Selenium can do it for you?
Dad duty1
  • Math TutorBedrock-generated practice problems and hints for a demanding 10-year-old.
Home lab and networking5
  • k3s clusterHome Kubernetes platform with a GPU worker: bare-metal load balancing, ingress, automated TLS, split-horizon DNS, encrypted secrets and Flux GitOps.
  • netwhyRust network-diagnostics CLI that tests interface, routing, DNS, TCP and TLS in order and stops at the first broken layer.
  • cartographerMaps and diffs multi-account AWS network topology (VPCs, Transit Gateway, Direct Connect, peering) without sending data anywhere.
  • neolinkA GStreamer-free RTSP backend in Rust for an open-source camera proxy.
  • Fedora workstation builderTired of rebuilding Fedora workstations by hand, I put myself out of business with Ansible. CI-tested on current Fedora.
Agent tooling4
  • ship-itAgent skill that walks work through stage gates with adversarial review, because an agent shouldn't grade its own homework.
  • VizCharterLocal MCP server that lets agents draw architecture diagrams and charts, with 1,100+ cloud icons.
  • Breaking Barriers hackathonCompeted with Team 46 in DC, building a sandbox that observes, summarizes and scores an agentic geospatial workflow end to end.
  • greplessOffline code graph for coding agents: 33 ms median reindex on a 3.2M-line repo.

When I'm away from a keyboard: trail running, sci-fi and history. English (fluent) and Spanish (native).