Baseer Builder
Nabeh's no-code computer-vision platform: collect data, label it, train models, run notebooks, deploy inference to live cameras and analyse the results. It's a team product with many people building it. The modules here are the ones I designed, built and own, along with the infrastructure the whole platform runs on.
largest deploymentEPM: Eastern Province Municipality, KSAA government project with 200+ use cases on 2,000+ cameras, sized for 50+ A100-class machines, delivered on Baseer Builder.
warehouse/training/k8s Job · on demand
Model Training
55h → 5htraining time with PyTorch DDP
- Training service for detection, classification and segmentation. Each run is an on-demand Kubernetes Job, so it survives interruptions, and Kubeflow Trainer keeps it going across reboots
- Distributed training with PyTorch DDP pools GPUs across machines and cut a production run from 55 hours to 5
2 more
- Live training metrics in MLflow, plus progress and custom statuses streamed over Kafka by a logger and a status service
- Pause, stop, resume and prioritise runs; pick the architecture and pretrained weights to trade speed for accuracy
deploy/deploy manager · k8s Jobs · Ansible
Deploy & Device Onboarding
5 days → 2 hrsto deploy the whole platform
- Deploys the entire Baseer Builder platform onto customer machines in about 2 hours, down from 5 days
- Zero-touch onboarding: nothing to install by hand. The deploy manager prepares each machine (Docker, NVIDIA drivers, container toolkit), creates or joins the Kubernetes cluster, and pre-pulls images
2 more
- Single and bulk onboarding, patching a device to a pinned version, and full cleanup on deboarding, all run as on-demand Kubernetes Jobs driving Ansible
- A 7-phase lifecycle reported to the backend over Kafka, with device health from kube-prometheus-stack on a unified dashboard
warehouse/notebooks/StatefulSet
Jupyter Notebooks
- Each notebook gets its own subdomain through ingress + MetalLB on bare metal
2 more
- Code persisted on NFS-backed NAS
- Idle detection warns the user, then shuts the notebook down to free GPUs
deploy/cameras/persistent service
Camera Service
- RTSP → HLS through MediaMTX with an API for the backend to add or remove streams
1 more
- Tuned for thousands of cameras with parallel conversion
warehouse/faces/k8s Job · on demand
Face Registration
- Supports multiple face models and writes embeddings to Qdrant
platform/infra/the foundation
Cluster & Delivery
- HA kubeadm cluster (HAProxy) across dev, prod and EPM
3 more
- Jenkins CI → image tags in an infra repo → Argo CD deploys
- Harbor registry on NAS used by 8 teams; Vault for every secret
- Moved InfluxDB to ClickHouse and every Docker-based service into k8s