Next-Generation Edge Intelligence & Agentic Workflows

01 — RUNTIME & ORCHESTRATION

Agentic orchestration

LangGraph + custom-built supervisors handle multi-step reasoning across our self-hosted LLM fleet (vLLM + LoRA fine-tunes on AMD MI50, fallback OpenRouter). Each agent runs sandboxed with explicit tool budgets, audit logs, and per-tenant isolation.

LangGraph vLLM LoRA OpenRouter AMD MI50

02 — EDGE COMPUTE

Edge & sovereign cloud

Inference-class workloads sit on the AWS Wavelength Zone in Casablanca — sovereign by Article 13 of Law 09-08 / CNDP. NVIDIA Jetson Orin and AMD edge boxes for in-store latency under 10 ms (POS correlation, real-time CCTV analytics).

AWS Wavelength Wavelength Casablanca NVIDIA Jetson Orin Cloudflare Edge

03 — DATA & STREAMS

Real-time stream processing

RTSP/ONVIF ingest from 11+ camera brands (Hikvision, Dahua, Axis, Hanwha…) → in-edge ONNX/TensorRT models → event bus → correlation with POS transactions. Backbone: Postgres (multi-tenant), object storage S3-compatible, Redis for hot state.

RTSP / ONVIF ONNX / TensorRT PostgreSQL Redis S3-compatible

04 — APPLICATION LAYER

Application & UI

Rust + Tauri for native desktop terminals (offline-first, HMAC-SHA256 transactional integrity). Flutter for in-store mobile (cashier handhelds, manager apps). Next.js + Cloudflare Workers for back-office and customer portals.

Rust Tauri Flutter Next.js Cloudflare Workers

05 — INFRA & OPS

Infrastructure & operations

Kubernetes (k3s for edge, EKS for cloud) provisioned via Terraform. CI/CD on GitHub Actions. Observability: Sentry, Prometheus, Loki. Secrets in HashiCorp Vault. Compliance trail: WORM-style audit log per tenant, exportable for CNDP / DGI inspections.

Kubernetes (k3s + EKS) Terraform GitHub Actions Sentry · Prometheus · Loki HashiCorp Vault

Back to home

How we pick a technical building block

A technical choice is judged over five years, not over a demo. We apply four criteria: project and community maturity, the real operating cost once in production, reversibility — can we leave without rewriting the application — and the ability to self-host if compliance demands it.

This filter rules out many attractive but closed tools. It also explains why our stack systematically favours open components: not out of ideology, but as a sovereignty and maintenance constraint.

The layers of our stack

Models and inference

We deploy open-weights models served by vLLM, locally or on Moroccan infrastructure. The issue is not raw model quality but cost per request at real volume: quantisation, context size, caching and routing to the smallest model that suffices.

Orchestration and agents

LangGraph for flows requiring explicit state and error recovery; n8n where readability by a non-developer team matters more. Both run on your infrastructure, so a business process never depends on a third-party service.

Edge and embedded

NVIDIA Jetson and RK3588 for vision inference at the edge. Working under hardware constraints enforces useful discipline: optimise the pipeline before adding compute, and measure end-to-end latency rather than model latency alone.

Application and interfaces

React and Vite for web, Go for high-concurrency services, Rust and Tauri when a lightweight installed client is required. The criterion is three-year maintenance load, not novelty.

Data and streams

PostgreSQL for transactional data, Redis for hot state, RTSP and ONVIF for video ingestion. Standard-protocol interoperability is what lets us plug into already-installed hardware rather than forcing a replacement.