Research & open data

Interactive quantitative analysis of the 15 Mexican presidential administrations (1934-2024). 9 structural variables (GDP, poverty, employment, violence, Gini, tax, investment, inflation, debt), LatAm-5 regional Z-score, Contextual Difficulty Index and 1,000-simulation bootstrap.

Practical case study of implementing a financial fraud detection system with ML for a LATAM fintech. Architecture, features, model, and production results.

Technical guide for migrating from a monolithic architecture to microservices. Patterns, anti-patterns, domain decisions, and lessons learned in production.

Analysis of the current state of AI adoption in Latin American enterprises. Trends, barriers, success stories, and opportunities by sector.

Fixed regulatory deadline: August 2, 2026. Fines up to €35M or 7% of global revenue. 20-point checklist, self-hosted stack for data residency, and remediation calendar for Latin American SaaS and fintech handling European data.

Production Docker Compose with Qdrant, Langfuse, LiteLLM Proxy, Redis 8, and Ollama on a $40 droplet. Nginx + Certbot, backups to Spaces, Grafana and Prometheus monitoring. Entire stack on a single server.

Spanish-language voice agent for dental clinics and small law firms. OSS stack: Pipecat for orchestration, Deepgram/Whisper.cpp for STT, Cartesia/Coqui TTS for voice, Qdrant for business RAG, Langfuse for traces, Redis for context.

Ten production-ready MCP servers: CRM, WhatsApp Business, CFDI invoicing, calendar, Postgres with RLS, and Qdrant RAG. Open repo numoru/mcp-templates-es with tests, OAuth, Docker, and Cloudflare Workers deployment.

Generative Engine Optimization explained with our own benchmark: 20 Mexican brands measured across ChatGPT, Perplexity, Gemini, and Claude. Measurement pipeline with Firecrawl, LiteLLM, and Qdrant.

Go script that runs 100 queries against 5 LLMs via LiteLLM proxy, vectorizes responses in Qdrant, and detects semantic citations. React dashboard. Published as numoru/geo-audit.

Practical guide to implementing a Model Context Protocol server in Go using mcp-go, integrating Postgres (row-level security), Gmail API, and Google Calendar. Includes OAuth, tests, and Cloudflare Workers deployment.

Tiered memory pattern for LLM agents: Redis for working memory (1h TTL), Langfuse sessions for auditable history, Mem0 for long-term semantic memory. Go code, when to expire, summarize, or vectorize.

Results of our own survey of 87 Latin American enterprises: frameworks used (LangGraph, CrewAI, Mastra), vector DBs (Qdrant, pgvector, Weaviate), observability (Langfuse, Phoenix), and real barriers. Only 23% report clear ROI.

From naive RAG to production RAG: Chonkie for semantic chunking, Qdrant with hybrid search (dense + BM25), self-hosted BGE-reranker, Anthropic Contextual Retrieval, RAPTOR, RedisVL semantic cache, and Ragas + Langfuse evaluation.

Full case: appointments agent + reminders agent + reviews agent, orchestrated with LangGraph and Postgres checkpointer. Mem0 semantic memory, Temporal long-running flows, healthcare guardrails, Langfuse traces.

Pipeline running the full eval suite on every PR: golden dataset versioned in Qdrant, automatic regression detection, RAG metrics with DeepEval, Langfuse traces, and merge blocking on score drops.

Cross of Fed, Deloitte, and local observations: which vertical pays most for AI, which is underdigitized, what stack applies in each (health → on-prem Ollama, legal → BAML, retail → Mem0 + Redis). 2x2 entry and ticket matrix.

How to audit AI code before production: Semgrep + Bearer + Trivy in pre-commit, SonarQube Community with AI-specific rules, DSPy for automatic test generation, Promptfoo for prompt validation in CI.

Fine-tuning Llama 3.3 8B on the Spanish ICD-11 catalog: Unsloth for efficient training, vLLM for serving, Qdrant for complementary RAG, lm-eval-harness benchmark. Model published on Hugging Face.

Exportable self-hosted n8n flow: semantic property search with Qdrant, description generation with Ollama (zero cost per listing), WhatsApp Business Cloud API, Langfuse traces. Repo numoru/n8n-blueprints.