Senior Backend Engineer focused on distributed systems, data-intensive Python/Go services, and AI/agent infrastructure on AWS.
I build backend systems where correctness, throughput, cost, and observability are measured rather than assumed. At work I'm accountable for data-extraction and processing platforms handling 120M+ requests/day; outside work I maintain a set of engineering labs covering infrastructure that AI teams need: agent evaluation, trace observability, rate limiting, async ingestion, and reproducible performance benchmarking.
| Area | Work |
|---|---|
| Backend systems | 5+ years building backend services, primarily in Python with Go for performance-critical tooling, including high-traffic data-extraction and processing platforms. |
| Distributed systems | Event-driven pipelines, async orchestration, Redis Lua atomicity, bounded concurrency, blue-green deployments, failure-aware data processing. |
| Cloud & cost engineering | AWS infrastructure redesign that cut compute spend by 50%, deployment pipeline time by 70%, and RDS operating cost by 25%. |
| AI / ML infrastructure | ML-driven workload optimization in production; agent evaluation, LLM-judge scoring, and trace observability tooling in the projects below. |
Core stack: Python, Go, PostgreSQL, Redis, MongoDB, FastAPI, Django, asyncio, AWS, Terraform, Docker, LLM APIs.
- Cloud migration: Architected the AWS Lambda → ECS Fargate migration for a 120M+ req/day scraping suite: load-tested at scale, cut over blue-green with zero downtime, removed Lambda's timeout and concurrency ceilings, and cut infrastructure cost by 50%.
- Database operations: Planned and executed a controlled PostgreSQL RDS major-version upgrade (15 → 18) for a 2M+ records/day workload, benchmarking performance before and after cutover: 40% faster analytical queries and 25% lower RDS cost through query-plan tuning and instance right-sizing.
- Data infrastructure: Designed a serverless event-driven ETL pipeline (Kinesis → Glue → S3 → Athena) that turns millions of daily JSON error events into queryable Parquet data-lake partitions, replacing ad-hoc log digging with real-time error analysis.
- ML in production: Deployed an SGD-regressor model that prioritizes scraping payloads by predicted value, lifting extraction rates by 20%, plus a reinforcement-learning multi-armed bandit that automated performance tuning across millions of daily database transactions.
- Leadership: Promoted to Senior after 16 months at mid-level for rebuilding the team through 60% turnover: onboarded a new engineering manager and two developers, grew the team back from 2 to 5 engineers, and standardized documentation templates and repository structure.
Event-driven benchmark platform for comparing AI code-review agents against versioned, reproducible PR-review cases.
- Fixes the inputs and stores raw outputs, so past runs can be re-scored under new rubrics without paying for new provider calls, solving the drift that makes most agent benchmarks incomparable.
- Redis Streams orchestration with explicit orchestrator/dispatcher/scoring/aggregator boundaries, MongoDB report persistence, FastAPI API, Typer CLI.
- Deterministic finding-match scoring plus configurable LLM-judge scoring across OpenAI and Anthropic adapters; sample run compares Claude and GPT agents with average score, P50/P95 latency, and failure-rate reporting.
- 462 tests,
mypy --strict, ruff, CI with secret scanning, seeded fixtures.
Django/PostgreSQL observability backend for LLM-agent traces: span ingestion, run finalization, metrics, and async quality evaluation.
- Moves agent debugging from console logs into queryable execution traces with cost, timing, errors, and evaluation lifecycle state.
- Idempotent ingestion, immutable completed-run semantics, hierarchical trace reconstruction, and database-level integrity constraints.
- Framework-agnostic fail-open Python tracer, durable persistence of async evaluator failures, Prometheus-style metrics.
Distributed rate-limiting library implementing token bucket, leaky bucket, and sliding-window log over local and Redis backends.
- Pushes quota decisions into Redis Lua scripts for atomic check-and-update, closing the race condition that breaks naive check-then-increment limiters across instances.
- Pure-function algorithm core with swappable backends, contract tests verifying local-vs-Redis behavioral equivalence, FastAPI/ASGI integration, configurable fail-open/fail-closed policy.
- Benchmarked at 197K–264K RPS in-process with microsecond overhead; a multi-instance simulation admitted exactly 100/100 requests under a global limit.
Reproducible, benchmark-backed experiments in throughput and memory behavior:
- Async Patterns: bounded-concurrency HTTP ingestion in Python: sync 130 RPS → async 2,500 RPS (20x), with circuit breakers, retry budgets, connection pooling, and backpressure.
- SQL Throughput Challenge: PostgreSQL bulk-read strategies compared under Docker resource constraints: multiprocessing at 124K rows/sec (5.3x baseline); async streaming with ~97% lower peak memory at 1M rows.
- Go Books Scraper: bounded-memory scraping ETL in Go: 2.2M items/sec in-memory pipeline at 4 allocs/op (108 items/sec end-to-end over live HTTP, rate limits respected), constant memory regardless of crawl size, Prometheus metrics.
| Category | Tools |
|---|---|
| Languages | Python, Go, SQL, Bash |
| Backend | FastAPI, Django, Django REST Framework, Pydantic, asyncio, aiohttp, asyncpg, REST API design |
| Data & Storage | PostgreSQL, Redis, MongoDB, Kinesis, S3 data lakes (Glue, Athena, Parquet) |
| AI / ML | LLM integration (OpenAI, Anthropic), agent evaluation, LLM-judge scoring, scikit-learn, multi-armed bandits |
| Cloud & DevOps | AWS (ECS, Lambda, RDS, CloudWatch), Terraform, Docker, GitHub Actions, Bitbucket Pipelines |
| Quality & Observability | pytest, mypy strict, ruff, Gitleaks, Prometheus metrics, structured logging, trace-style telemetry |
- AWS Certified Machine Learning Engineer – Associate
- AWS Certified Developer – Associate
- Anthropic: MCP: Advanced Topics and Claude Code in Action
Happy to talk about distributed systems, database performance, cost engineering, and AI agent infrastructure.
- Email: alumlira@gmail.com
- LinkedIn: linkedin.com/in/aluiziolira
- Location: Recife, Brazil | Remote