Loading content…
QBS Co.
AI-orchestration app reusing a single distributed CV pipeline to power multi-domain surveillance products
.NET CoreC#PythonGoC++FastAPIPyTorchTensorRTONNXgRPCREST APIsKubernetesMinIO SDKMicroservicesMicrofrontendsWebpack Module FederationTemporal.ioByteTrackDeepSORTOC-SORTBoT-SORT
VERSEYE is an enterprise AI app market orchestration platform engineered from scratch at QBS Co., where Noman Ali serves as Technical Lead — directing a cross-functional team across computer vision, AI/ML, backend, frontend, DevOps, and analytics.
Platform Overview The core idea: one high-throughput computer vision pipeline, many domain-specific surveillance products. VERSEYE powers employee monitoring, restaurant monitoring, planogram compliance, PPE monitoring, footfall analytics, and industrial quality inspection — all from a single shared pipeline backbone with a distributed backend and micro-frontend UI.
Dynamic Plugin Runtime (.NET + Kubernetes) Plugins are containerized AI modules deployed, scaled, and managed at runtime directly from the UI. Using .NET with native kubectl integration, operators can launch or retire detection models on live Kubernetes clusters with zero downtime — no DevOps ticket required.
Polyglot Pipeline Engine The core pipeline is Python-based (FastAPI + gRPC), but plugins can be written in any language — C++, Go, Python, or C# — as long as they expose a gRPC or REST interface. This enables teams to pick the best runtime per model: CUDA-optimized C++ for object detection, Go for high-throughput ingestion, Python for model experimentation.
Connected Evidence Graph & Natural Language Search Because every detection in the pipeline is linked: a natural language query like "give me all people who are using mobile" resolves through text → local track ID → global track ID → frame snapshots → timestamp-synchronized video clips, all stored across distributed MinIO object storage. No manual tagging. No keyword search. Pure semantic retrieval.
Multi-Camera Tracking & Global ReID Supports ByteTrack, DeepSORT, OC-SORT, and BoT-SORT for multi-camera tracking with global re-identification across face, gait, pose, appearance, and text query signals — enabling continuous subject tracking across camera boundaries in large-scale deployments.
Micro-Frontend Shell (Webpack Module Federation) Each surveillance domain (PPE, footfall, employee monitoring, etc.) is an independently built and deployed micro-frontend. Webpack Module Federation dynamically loads these remotes into a unified shell, allowing domain teams to ship UI updates without coordinating a monolith release.
Inference Optimization & Durable Retraining Models are optimized end-to-end: ONNX export → genetic-algorithm-based graph pruning → TensorRT FP16/INT8 acceleration on NVIDIA GPUs. Temporal.io orchestrates durable retraining workflows, ensuring model lifecycle jobs survive hardware failures and are fully auditable.
The platform has been demonstrated in successful client demos and is under active production feature development.
Distributed AI orchestration and computer vision pipeline combining a .NET kubectl Kubernetes operator, Python/FastAPI CV engine, polyglot gRPC plugin runtime, MinIO distributed object storage, and a Webpack Module Federation micro-frontend shell.
Micro-Frontend Shell & Domain Remotes
Webpack Module Federation host shell dynamically loading isolated per-plugin domain frontends (PPE, Footfall, Employee, Planogram, Quality).
Webpack Module FederationReactTypeScript
Runtime Plugin Orchestration Layer
.NET service managing dynamic container lifecycle, plugin registry, and automated scaling via kubectl commands on Kubernetes clusters.
.NET CoreC#KubectlKubernetesDocker
Core CV Pipeline & Multi-Camera Tracking Engine
High-performance Python engine executing chained inference pipelines (ByteTrack, DeepSORT, OC-SORT, BoT-SORT), global ReID, and semantic text-to-visual search.
PythonFastAPIPyTorchByteTrackBoT-SORTDeepSORTOC-SORT
Polyglot Plugin Workloads
Extensible detection and classification plugins written in C++, Go, or Python communicating with the core engine over gRPC and REST.
C++GoPythongRPCREST APIs
Inference Optimization & Durable Workflows
Temporal.io orchestration for durable retraining pipelines, genetic algorithm model pruning, ONNX Runtime, and TensorRT GPU execution.
Temporal.ioTensorRTONNXCUDA
Distributed Storage & Connected Evidence Layer
MinIO S3-compatible distributed object storage persisting indexed frame snapshots, timestamped video segments, and relational track metadata in PostgreSQL and Redis.
MinIO SDKPostgreSQLRedisDistributed Storage
Skills
