Israel Adolfo López Maíz

Celium is a distributed platform for deploying and running AI models across cloud, on-prem, and edge environments. The system operates over heterogeneous GPU-enabled nodes with strict latency, availability, and cost constraints. Action

• • Designed and optimized the end-to-end model deployment lifecycle.

• • Implemented orchestration, monitoring, and recovery modules in Python and Rust.

• • Integrated C-based execution paths for latency-critical operations.

• • Automated build, deployment, and rollback pipelines using Docker, Helm, and CI/CD.

• • Tuned CPU/GPU scheduling, affinity, and resource allocation policies. Metrics

• • Reduced model time-to-production from hours to minutes.

• • Improved sustained GPU utilization by ~20–30%.

• • Lowered P95 inference latency by ~30–40% on critical paths.

• • Reduced deployment-related incidents by ~40%. Result Enabled fast, reliable, and cost-efficient AI model deployment at scale. Improved platform stability and

predictability while supporting decentralized execution and continued network growth.

Context Mission-critical defense platform developed under the European Defence Fund. The system targets next

generation military vehicles and must operate under strict availability, security, and interoperability constraints across NATO environments. Action

• • Designed and implemented backend services in Java and Node.js using Clean Architecture.

• • Integrated a RAG system for low-latency contextual retrieval of operational intelligence.

• • Developed GIS components for real-time geospatial analysis and tactical visualization.

• • Defined secure APIs and interfaces to interoperate with NATO-compliant systems.

• • Automated CI/CD pipelines and Kubernetes deployments with full observability. Metrics

• • Reduced operational data retrieval time from seconds to sub-second latency.

• • Improved service availability to >99.9% in simulated hostile environments.

• • Decreased deployment and integration errors by ~35% via automated CI/CD.

Cloud-native security analytics platform ingesting massive volumes of sensitive logs. The ingestion layer must handle bursty traffic, strict reliability targets, and compliance constraints (FedRAMP/FIPS/NIST). Action

• • Owned and optimized the ingestion load balancer for sustained high-throughput production traffic.

• • Improved fault tolerance, recovery paths, and observability across ingestion services.

• • Implemented reliability features (including RELP support) to reduce event loss under failure scenarios.

• • Built multi-region / multi-cloud redundancy mechanisms to improve disaster recovery posture.

• • Integrated quality and security controls into CI/CD to support FedRAMP readiness. Metrics

• • Sustained ~60 TB/day ingestion under production load.

• • Reduced event loss and improved delivery guarantees during network/service failures.

• • Lowered incident frequency and improved detection time via stronger telemetry and alerting.

• • Improved recovery time for ingestion components through hardened failover and runbooks. Result Delivered a faster and more reliable ingestion layer for mission-critical security data. Increased

operational resilience and strengthened the compliance path while keeping performance stable at scale.