BCloud Solutions | RAG Systems & MLOps for SaaS (@BCloudToday) header
BCloud Solutions | RAG Systems & MLOps for SaaS (@BCloudToday) profile picture

Helping SaaS startups build RAG systems & AI infrastructure at scale | AWS/Azure certified | Built production AI for MasterSuiteAI & VoxRoute | Spain 🇪🇸

Barcelona
Joined December 2022
Photo by BCloud Solutions | RAG Systems & MLOps for SaaS (@BCloudToday)Photo by BCloud Solutions | RAG Systems & MLOps for SaaS (@BCloudToday)Photo by BCloud Solutions | RAG Systems & MLOps for SaaS (@BCloudToday)Photo by BCloud Solutions | RAG Systems & MLOps for SaaS (@BCloudToday)Photo by BCloud Solutions | RAG Systems & MLOps for SaaS (@BCloudToday)Photo by BCloud Solutions | RAG Systems & MLOps for SaaS (@BCloudToday)Photo by BCloud Solutions | RAG Systems & MLOps for SaaS (@BCloudToday)Photo by BCloud Solutions | RAG Systems & MLOps for SaaS (@BCloudToday)Photo by BCloud Solutions | RAG Systems & MLOps for SaaS (@BCloudToday)Photo by BCloud Solutions | RAG Systems & MLOps for SaaS (@BCloudToday)
BCloud Solutions | RAG Systems & MLOps for SaaS retweeted
I am thrilled about how we are integrating AI and LLM models into our daily lives. This video was generated by NotebookLLM (a fascinating development). youtube.com/watch?v=Pt8RuBUf… any attribution for @BCloudToday
1
1
116
Tu LLM promete 1M tokens. En produccion rinde como 100K. Context Rot: degradacion catastrofica procesando contextos largos. Tu modelo 200K tokens se vuelve poco confiable alrededor de 130K. Latencia 500ms → 60 segundos. Y nadie te aviso.
1
76
Soluciones practicas: 1. Calcula tu limite REAL (70-80% del anunciado) 2. Implementa chunking semantico (no por tokens) 3. Usa re-ranking antes de inyectar contexto 4. Monitorea accuracy por longitud de contexto 5. Considera modelos especializados para tu caso
1
23
4676% de facturacion extra GCP. De 60K esperados a 2M reales. Thread con los 5 errores que causan billing shocks y el framework FinOps de 7 pasos para evitarlos:
1
13
Implementacion rapida de alertas GCP: gcloud billing budgets create \ --billing-account=ACCOUNT_ID \ --display-name="Alerta 50%" \ --budget-amount=1000 \ --threshold-rule=percent=0.5 Esto toma 5 minutos. Y puede salvarte miles.
1
10
82% de empresas adoptan agentes IA. La mayoria falla en monitorearlos. 40% de proyectos cancelados para 2027 por falta de observabilidad (Gartner). Thread sobre como no ser parte del 40%:
1
10
Herramientas comparadas: Langfuse: + Open source, self-host posible + Buena integracion LangChain LangSmith: + Mejor UX + Parte del ecosistema LangChain - Solo cloud W&B: + Mejor para experimentos - Mas orientado a training que inference
1
16
Checklist seguridad inmediata: 1. OAuth 2.1 para auth 2. Sanitizar TODOS los inputs 3. Allowlist comandos (no blocklist) 4. Logging completo 5. Rate limiting 6. Container aislado MCP El #3 critico. Blocklists siempre fallan. Thread 4/6
1
30
Ejemplo sanitizacion: ALLOWED = {"list_files", "read"} def execute_tool(name, args): if name not in ALLOWED: raise SecurityError() validated = validate_args(name, args) return tools[name](**validated) Thread 5/6
16