AI coverage prediction on a radio digital twin
A detailed radio simulation shows where an antenna's signal will reach, but slowly. This system trains a small AI model on thousands of simulated coverage maps, so it predicts new ones almost instantly and helps choose the best antenna tilt and power.
- NVIDIA Sionna RT
- PyTorch
- NVIDIA CUDA
Telecom MLOps with drift-triggered retraining
Models on network data go stale as customers, traffic and equipment change. This pipeline runs six telecom predictions, from churn to fault root cause, through one daily loop that spots change, retrains, and only switches to a new model when it is clearly better.
- MLflow
- Evidently AI
- XGBoost
- scikit-learn
Medallion lakehouse for RAN KPIs
Network performance data arrives as vendor files every 15 minutes, often late, duplicated or renamed. This lakehouse turns them into one trusted, versioned set of KPIs, served through an API for reporting, planning and what-if analysis, on a simulated 4G and 2G network.
- Apache Iceberg
- DuckDB
- FastAPI
- Google OR-Tools
Agentic AI for NOC incident triage
A NOC operator often checks several systems by hand to find why calls drop. Here, four AI agents read the complaint, check network events and call records, and write an incident ticket with a recommended action, in one step.
- Google Gemini
- Google ADK
- Vertex AI
- Model Context Protocol
Edge SLM for MCP tool calling
Most AI assistants carry a long tool manual in every request. This very small model knows its 14 tools by heart, so a short request becomes the right action straight away, on a laptop with no cloud and no per-use cost.
- Google Gemma 3
- LoRA
- Anthropic Claude
- Ollama
Scale-to-zero GPU inference on Kubernetes
An idle GPU for an in-house AI model still costs money. This platform switches the GPU on when requests arrive and off when the queue is empty, keeps the first answer quick, and holds requests safely in line if a GPU machine drops out.
- Google Kubernetes Engine
- KEDA
- vLLM
- NVIDIA GPU
Built, measured and public.
Six systems for telecom data, ML and AI, from radio coverage prediction and drift-triggered MLOps to a RAN KPI lakehouse, agentic NOC triage, an edge tool-calling model and scale-to-zero GPU inference. Each one is measured and its code is public.
Built and published by our AI lead.