sergio.dev
LinkedIn Get in Touch
Facet · Infra

Platform & LLM Infra

The foundation that makes everything else look easy.

Agents impress in a demo; platforms hold up in production. The part of the stack nobody tweets about is the part that decides whether the system survives: the gateway that unifies providers with metering and fallback, the data reconciled to the cent before cutover, the deploy that happens without me at the keyboard. I build this layer first — and that's why I can ship everything else fast.

Multi-LLM Gateway OpenAI-Compatible API SSE Streaming Metering & Billing PostgreSQL · pgvector ETL & Data Migration Kafka · RabbitMQ Auto-Deploy (DigitalOcean) Query Tuning Cost Telemetry Multi-LLM Gateway OpenAI-Compatible API SSE Streaming Metering & Billing PostgreSQL · pgvector ETL & Data Migration Kafka · RabbitMQ Auto-Deploy (DigitalOcean) Query Tuning Cost Telemetry
Talk to me 📄 CV — Agentic Solutions Architect (PDF)
0 Providers behind a single API
0 Records/day processed in batch
~99.9% Multi-year uptime in production
0 Query-time cut at Bradesco

The center of my platform is Railter: an OpenAI-compatible gateway that unifies Claude, GPT, and Gemini behind a single API, with per-key auth, metering, billing, and retry/fallback. None of my products talks to a provider directly — when a new app is born, it's born with cost observability, SSE streaming, and fallback for free. Switching models is config, not a refactor.

LLM economics is an infrastructure discipline, not an end-of-month optimization. Tier routing (GLM → Sonnet → Opus) picks the cheapest model that holds quality, with an escalation lock in code and per-worker-session cost telemetry. The eval harness is part of the decision: a tier only changes when the data shows it has to.

Underneath that, the enterprise ballast: at Monument, I migrated dozens of clients with idempotent pipelines reconciled to the cent; at YOUSE, I integrated microservices into a 50+ service Kafka mesh; at Bradesco, I processed 1M records/day in batch with tuning that cut query time by 60%; at DETRAN, I held 99.9% uptime for years. Today that repertoire runs a fleet of Rails apps on DigitalOcean — auto-deploy with migrations, a shared Postgres cluster, and the habit of sizing connection pools before they become an incident.

Architecture

The backbone: every app talks to Railter — one API, three providers — and lives on the same foundation of Postgres, queues, and auto-deploy on DigitalOcean.

Projects

Through this lens

Infra · LLM Gateway Production

Railter

Internal LLM gateway. OpenAI-compatible, SSE streaming, per-key usage & billing.

Seen as a platform, Railter is the piece that pays the rent: a single API contract for the whole ecosystem, with cost, streaming, and fallback solved exactly once. A new product doesn't pick an SDK, doesn't manage provider keys, doesn't reimplement retry — it points at the gateway and inherits the entire infrastructure.

Impact
  • One OpenAI-compatible API in front of Claude, GPT, and Gemini
  • Per-key metering and billing — cost visible per product
  • Retry, fallback, and SSE solved in the gateway, not in every app
Stack
Rails 8 OpenAI-compatible API SSE Streaming Multi-LLM Routing Usage Metering
Economy · Multi-LLM Production · Private

Economy Routing

Cost routing across model tiers with an escalation lock and an eval harness.

Through the infra lens, economy routing is capacity planning for intelligence: which model tier serves each class of task at the lowest cost, with quality proven by an eval harness and an escalation lock in code. It's the difference between an LLM bill that scares you at month's end and predictable cost per session.

Impact
  • Tier routing (GLM → Sonnet → Opus) with an escalation lock
  • Per-worker-session cost measured, not estimated
  • The eval harness decides tier changes with data, not faith
Stack
OpenRouter Multi-LLM (GLM · Claude) Eval Harness Ruby on Rails 8
Also in: Harness
Data · Migration Delivered

Monument Migration Platform

Enterprise migrations reconciled to the cent, with Claude Code as the dev driver.

Enterprise migration is data infrastructure under pressure: heterogeneous schemas from competitor platforms, real financial math — interest, fees, amortization — and a single acceptance criterion: match to the cent before cutover. Idempotent pipelines with multi-level validation turned that into routine, not heroics.

Impact
  • Idempotent pipelines: running again never makes the state worse
  • Multi-level validation: unit, customer, facility, period
  • To-the-cent financial reconciliation as the cutover gate
Stack
TypeScript / Node.js MySQL AWS Claude Code
Flagship · Multi-Agent · MCP Production · Private

Bailder

Personal multi-agent orchestrator. WhatsApp-native, autonomous Claude Code workers, tier-enforced safety.

Under the infra lens, Bailder is a fleet: new-project bootstrap (repo → DB → DNS → deploy) in 1 tool call, auto-deploy that runs migrations on its own after the health check, and multiple apps living on the same Postgres cluster — with the scars turned into rules, like sizing connection pools before the incident.

Impact
  • Full new-app bootstrap in 1 tool call
  • Auto-deploy with automatic migrations after the health check
  • Multi-app on a shared Postgres cluster, pools sized right
Stack
Ruby on Rails 8 Claude Code SDK MCP (Playwright) Multi-LLM (OpenRouter) Solid Queue WhatsApp Cloud API DigitalOcean (auto-deploy)
Also in: Harness
Experience

The track record, re-read

Aug 2025 – May 2026

Senior Data Migration Engineer · AI-Augmented

Monument (Remote, US)

  • Dozens of enterprise clients migrated with to-the-cent financial reconciliation
  • Idempotent pipelines + multi-level validation as the cutover quality contract
Claude Code TypeScript Node.js MySQL AWS
Nov 2022 – May 2023

Senior Software Engineer (Consultant)

META | YOUSE — Caixa Digital Insurance (Remote)

  • Microservices integrated into an event-driven Kafka mesh with 50+ services
  • Legacy data migration critical to risk assessment at the country's largest digital insurer
Node.js Ruby React AWS (S3, RDS, ECS) Kafka
Nov 2020 – Nov 2021

Java Software Developer

MJV Technology & Innovation | Bradesco Insurance (Remote)

  • Async batch of up to 1M records/day with ETL on Apache Airflow
  • PostgreSQL tuning (indexes, partitioning, queries) cutting average time by 60%
Java Spring Boot PostgreSQL Apache Airflow
Jan 2017 – Nov 2020

Java Developer

PSG | DETRAN — Government Traffic Dept. (Remote)

  • 99.9% multi-year uptime on a government traffic-violations system
  • Multi-million-vehicle scale integrated with multiple public APIs
Java Spring Boot PostgreSQL Heroku
Skills

The toolkit for this lens

📊

Data & Vector Stores

Schema design, query tuning, RAG over proprietary data. pgvector-first, lifted from production scars.

PostgreSQL (tuning · partitioning) pgvector / RAG BigQuery · Looker Elasticsearch MongoDB · MySQL ETL & Data Migration
☁

Cloud & DevOps

Opinionated infra. Auto-deploy from git, observability that survives outages, cost-aware multi-tenant patterns.

DigitalOcean (App Platform · DBs) AWS (S3, RDS, ECS, Lambda) Docker · Compose Kubernetes GitHub Actions · GitLab CI Systemd Auto-Deploy
🏗

Distributed Systems

Event-driven architectures, bounded contexts, async-first messaging. Scaled monoliths into services the right way.

Event-Driven Architecture Kafka · RabbitMQ Microservices · DDD CQRS / Event Sourcing System Design
⚙

Backend & APIs

Robust APIs and scalable systems with a focus on correctness, performance, and operational sanity.

Ruby on Rails 8 Python (FastAPI) Node.js / TypeScript REST · GraphQL · SSE Sidekiq · Solid Queue Background Jobs

Does your AI stack need a foundation?

Multi-LLM gateway, data at scale, and agentic architecture — as a platform role or as solution-architecture consulting.

Talk to me 📄 CV — Agentic Solutions Architect (PDF)
SB
Sérgio Brito
Online
👤 About 📚 Projects 💻 GitHub ✉ Get in Touch