Our Technology Stack

Component-based rendering engine for all operator dashboards, data-viz panels and configuration screens. We leverage Server Components for improved TTFB and streaming for progressive data delivery.

Full-stack React framework providing file-based routing, server actions, ISR/SSG, and Edge Runtime support. Powers the Sky Nexus web application and public-facing site.

Strict static typing across all frontend and shared packages eliminates entire classes of runtime errors and accelerates onboarding through self-documenting contracts.

Utility-first CSS framework with design-token integration, dark-mode support and a JIT engine. Enables consistent theming and rapid UI composition without bespoke stylesheets.

Accessible, unstyled primitives (dialogs, menus, popovers, tabs) combined with the shadcn component registry for a production-ready, customisable component library.

D3 handles custom, physics-based and topological visualisations (network graphs, digital twin maps). Recharts handles standard operational charts — line, area, bar, radial.

Declarative animation library used for dashboard transitions, alert animations and on-boarding flows — kept minimal to preserve performance budgets.

Server-state management with smart caching, background refetch and optimistic updates. Powers all live polling on operational dashboards where WebSockets are not appropriate.

Lightweight client-state store used for UI preferences, sidebar state, command palette and cross-component shared state without the boilerplate of Redux.

Vitest for unit and component tests with coverage; Playwright for end-to-end browser tests across Chromium, Firefox and WebKit in CI.

The backbone of all transactional data — tenants, users, workflow definitions, audit logs, configuration and billing. Leveraged with JSONB for semi-structured fields and pg_partman for time-based partitioning.

A PostgreSQL extension that turns time-series data into a first-class citizen. Stores all operational telemetry, metrics and KPI history with automated hypertable chunking and continuous aggregates.

Handles sub-second IoT sensor streams and infrastructure metrics where TimescaleDB write throughput is insufficient. Columnar storage with flight SQL enables fast range queries.

Used for the integration connector configuration store, dynamic form schemas, and event payloads where schema flexibility is paramount. Atlas Search powers full-text search across resources.

Multi-purpose: L2 cache for API responses, session store, distributed rate limiting, real-time pub/sub bus for dashboard events, and queuing with Redis Streams.

Column-oriented OLAP database powering all analytical queries — historical trend reports, cross-tenant aggregations and executive dashboards. Ingests billions of events with sub-second query latency.

On-premises and cloud-compatible object storage for audit log archives, ML training datasets, report exports, workflow attachments and container image layers.

Distributed full-text and vector search across operational logs, documentation, case studies and resource metadata. Used alongside the ELK stack for structured log aggregation.

Primary deep learning framework for anomaly detection, predictive maintenance, natural language understanding and time-series forecasting models. Used with FSDP for distributed training.

Handles classification, regression and clustering workloads where interpretability and training speed outweigh the benefit of deep learning — particularly in compliance risk scoring and fraud indicators.

Transformer models for document understanding, log summarisation and natural language workflow creation. Both hosted models and on-premises fine-tuned variants are supported.

Tracks every training run: parameters, metrics, artefacts and model versions. The model registry gates promotion from staging to production with mandatory evaluation thresholds.

Distributed execution framework for large-scale hyperparameter search, batch inference across millions of events, and data preprocessing pipelines that exceed single-node memory.

Models are exported to ONNX for hardware-accelerated inference at the edge and on CPU/GPU in production. ONNX Runtime delivers up to 4× latency improvement over Python-based serving.

Kubernetes-native ML pipeline orchestration for reproducible, parameterised training and evaluation workflows. Integrated with ArgoCD for GitOps-driven pipeline promotion.

Orchestration frameworks for retrieval-augmented generation (RAG), agentic workflows and document Q&A within the platform's AI Decision Support module.

Continuous monitoring of deployed models for data drift, concept drift and performance degradation. Triggers automatic retraining pipelines when models fall below quality thresholds.

Centralised feature store ensuring consistent feature engineering between training and inference. Prevents training/serving skew — one of the most common sources of ML production failures.

All build, test, lint, scan and container image pipelines run on GitHub Actions with reusable workflow templates across all repositories. Self-hosted runners handle GPU-intensive ML builds.

Declarative continuous delivery with automatic sync from Git, multi-cluster support, progressive delivery and SSO integration. Every production change is a pull request merge.

Canary and blue/green deployments with automated analysis gates. New versions receive a percentage of traffic while metrics are evaluated — rolled back automatically on degradation.

Static analysis for code quality, security hotspots, code smell and test coverage gates. Pull requests are blocked from merging when quality gates fail.

Automated dependency version management across all repositories. Renovate opens PRs for new versions, runs the full test suite and auto-merges patch updates on success.

Software Bill of Materials generated for every container image. SBOMs are signed with Cosign and published to our transparency log, enabling compliance with supply chain security mandates.

Reproducible, containerised build system that ensures 'works on my machine' becomes 'works everywhere'. Earthfiles replace complex Makefile chains across polyglot repositories.

All services emit traces, metrics and logs in the OpenTelemetry format. A single, vendor-neutral instrumentation layer means we can swap backends without touching application code.

Prometheus scrapes all service and infrastructure metrics. VictoriaMetrics provides long-term, cost-efficient storage with superior query performance at petabyte scale.

Unified observability UI with dashboards for infrastructure, services, SLOs and business KPIs. Grafana Alerting manages all alert rules with multi-channel notification routing.

End-to-end distributed tracing across all microservices. Tempo stores traces in object storage at low cost; Jaeger is used where native UI and adaptive sampling are required.

Horizontally scalable log aggregation without indexing. Promtail/Vector ship logs from all pods and nodes; LogQL enables fast ad-hoc log querying directly from Grafana.

Always-on CPU, memory and goroutine profiling for all Go services. Flame graphs integrated into Grafana dashboards enable root-cause analysis without profiling on demand.

External synthetic checks confirm customer-facing endpoints are reachable from multiple global probe locations. Feeds directly into public and private status pages.

Alert routing, on-call scheduling, escalation policies and post-incident review workflows. Deep integration with Slack, Jira and the Sky Nexus platform itself.

Production-grade Kubernetes in 40MB. Runs on edge hardware as low as 512MB RAM, enabling the same Kubernetes tooling and GitOps workflows at the edge as in the cloud.

MQTT brokers for device-to-platform connectivity. EMQX clusters handle millions of device connections in the cloud; Mosquitto runs as the local edge broker for offline-capable deployments.

Flow-based visual programming for sensor data transformation, protocol bridging (Modbus, OPC-UA, BACnet → MQTT) and local automation logic on edge devices.

Native support for industrial automation protocols. OPC-UA connects to modern PLCs and SCADA systems; Modbus bridges legacy industrial equipment into the digital operations layer.

For organisations already on Azure or AWS, Sky Nexus deploys as IoT Edge modules or Greengrass components, providing a native cloud-managed edge runtime with built-in OTA update support.

Lightweight, sandboxed Wasm modules run ML inference and data filtering at the edge without a container runtime. Enables sub-10ms inference on resource-constrained devices.