I build systems that put LLMs into production — from streaming data pipelines to eval infrastructure that catches prompt regressions before deploy.

Use Claude Code as an architecture amplifier: I define the system boundaries, own the decisions, and drive the execution.

LLM apps regress when prompts change without control. PromptBench treats prompts as versioned software assets.

• PromptVersion is immutable — no UPDATE, every edit creates an auditable snapshot
• Measures quality, latency, variance, and real cost (tokens from the provider's usage response, not estimates)
• Regression verdict: dimension ∧ slice ∧ cost — a better global average doesn't earn promotion if any slice regresses
• 87% test coverage on the evaluation module; deterministic checks are pure functions

Stack: Next.js · FastAPI · PostgreSQL 16 · Redis · arq · Anthropic Claude

Political opinion monitoring across 167 municipalities in Rio Grande do Norte, Brazil — at < R$5/month.

• RSS + social media ingestion → NLP sentiment pipeline → LLM-generated strategic recommendations
• Real-time dashboard with RAG over regional political knowledge base
• Cloud-native on GCP with full data governance

Stack: Python · GCP · BigQuery · dbt · Cloud Run

Exactly-once streaming pipeline with Claude as the anomaly explainer — ~US$40/month on GCP.

• Redpanda → ksqlDB → Apache Iceberg with exactly-once semantics
• Claude explains detected anomalies in natural language, not just flags them
• Cost-controlled LLM integration with token tracking

Stack: Redpanda · ksqlDB · Apache Iceberg · Python · Anthropic Claude

Every project here has:

• A problem statement — not just a stack demo
• ADRs (Architecture Decision Records) explaining why each key decision was made
• A build audit trail — docs/progress/blockN.md logging what was done, what ran, what was decided
• Real numbers — cost, latency, test coverage — not vague claims

AWS Certified Cloud Practitioner · Microsoft Azure AI Essentials · Anthropic Claude 101 · Anthropic Claude Code in Action

• Expanding PromptBench with multi-model judge (majority vote), continuous eval on real traffic, and GitHub Actions PR gate
• Growing English fluency for international collaboration