ai-ran

Research-grade Python framework for LLM-assisted RAN slicing and radio resource management in 5G O-RAN, accepted by INFOCOM Workshop 2026.

GPU-accelerated Network Digital Twin for AI-RAN experiments in Munich. Powered by NVIDIA Sionna RT and PyTorch for ray-traced RF simulation. Includes real-world Munich simulation, Mitsuba 3 backend, and finite-difference Adam optimization for antenna tuning. Automated KPI reporting (SINR/throughput) and coverage mapping for 6G R&D workflows

Official repository of "Large multimodal model-based environment-aware mobility management"

Single-node K3s systems study: measure latency (p50/p95/p99), jitter, and a throughput proxy when a vDU-style FFT and a GEMM inference workload share one NVIDIA GPU via the Kubernetes Device Plugin with CUDA time-slicing. Reproducible PyTorch workers, Job manifests, build/run orchestrator, and CSV/plots pipeline for edge AI-RAN.

Curated papers on AI-RAN, AI for RAN, Open RAN intelligence, and autonomous telecom network management.

Sovereign Speculative Tool Execution & Persistent Radix KV-Cache Engine for NVIDIA TensorRT-LLM, NIM & Blackwell Clusters with Cloudflare Clef, Hermes Paperclip, and Enterprise AI Control Plane

AI-RAN KPI forecasting system for RAN telemetry, operational intelligence, and network performance analysis.

A concept for turning carrier networks into distributed edge intelligence infrastructure. White paper on AI-RAN, cooperative perception, and network-native AI services, covering architecture, data rights, and governance. CC BY 4.0.

Technical portfolio accompanying AI-RAN Engineering: Intelligent Radio Networks for 5G and Future 6G by Sébastien Beyh, PhD.

Clean-room port of NVIDIA Aerial 5G L1 to AWS Trainium via NKI. Agentic, hardware-in-the-loop. Views are the author's own.

Report-aligned DVB-S2 broadcast satellite link sizing for East Africa with MATLAB ML models for MCS, EIRP, capacity, availability, link margin, and resource allocation.

A small, domain-specific language model for AI-RAN mobility management. It reads UE measurement reports as token sequences and generates a handover decision, a target cell and a short natural-language rationale. It is meant to run as a near-RT RIC xApp (or an AI-RAN inference service next to the gNB).

AegisRAN: failure-aware AI-RAN control-plane simulator for 6G edge systems, with counterfactual reward labeling and learned rerankers.

Federated learning for CSI feedback compression - CsiNet + FedAvg under non-IID channel statistics

CUDA-native MAC scheduler with cache-aware kernel design and dynamic SM partitioning for 6G AI-Ran edge nodes