Skip to content
7AZZANI 7AZZANI
Support Log in Sign up

7A / SIGNAL SAUDI ARABIA —:— PHILIPPINES —:— ENGLISH, ARABIC, TAGALOG

Azzani

Engineer, founder of 7AZZANI, builder of AI assistants and the systems around them.

I'm Abdullah Azzani. For more than fourteen years I've worked in programming and software development, and these days most of my time goes into AI assistants that talk to real customers, the platforms behind them, and 7AZZANI, where I share the tools and the lessons.

NowVoice and chat assistants that work in Arabic and English, and self-hosted AI systems businesses can own.

Talk to 7A See the systems
Abdullah Azzani
14+years building
7A · SCRIPTED · LOCAL TURN 00

    03 / Identity

    The short version.

    Based

    Saudi Arabia & Philippines

    —:— in Saudi Arabia · —:— in Philippines · English, Arabic, Tagalog

    Years in technology

    14+

    AI assistants

    8+

    4 shown here; the rest are private

    GitHub

    — public repos

    github.com/7AZZANI

    7AZZANI, live

    • 60published articles
    • 9free developer tools
    • 85community questions
    Open the platform

    Now building

    Voice and chat assistants that work in Arabic and English, and self-hosted AI systems businesses can own.

    Disciplines

    PythonDjangoFastAPIAPI engineeringBash & shell automationJavaScriptReactResponsive UXAI agentsRAG pipelinesVoice interfacesPyTorchHugging Face stackWorkflow automationNVIDIA GPU stackNVLink & MIGvLLM servingEmbeddings inferenceJupyterHubModel catalogueLLM gateway & meteringLinuxSlurmKubernetesDockerStorage & LVMnginx & TraefikPrometheusGPU telemetryCentralised loggingRunbooks & handoverKeycloakWireGuard & OpenVPNSecurity engineeringPostgreSQLRedis & CeleryVector searchData pipelines

    Not sure where to start?

    Ask 7A — it knows the whole page.

    Ask 7A

    04 / Systems

    What I've built, screen by screen.

    Each record says what the system is, what I did, and — where it matters — how it was built. The screens are the real ones.

    ATCL AI ASSISTANTATCL ecosystem view with the four divisions around the core
    ATCL ecosystem view with the four divisions around the core The ecosystem view where the assistant greets visitors
    Role
    Technology consultant and lead capture for enterprise and government buyers
    Languages
    English, Arabic (right-to-left)
    Channels
    Web chat, voice with barge-in, PDF hand-over, callback leads

    01 / AI assistantsLive

    ATCL AI Assistant

    Built end to end

    01The system

    A bilingual voice-and-chat consultant that is the front of house for a Saudi ICT company.

    On atcl.nzum.space the assistant greets visitors in English or Arabic, answers only from a curated knowledge base about ATCL’s infrastructure, AI, software, cybersecurity and managed services, hands over brochures, and captures callback requests as leads — by text, or by voice with barge-in.

    02What I did

    Designed and built it end to end: the persona and guardrails, the tool calls for document hand-over and lead capture, server-owned consent, the cost guards, and the ecosystem “constellation” view it lives in.

    03How I built it

    Measured before deciding. Spoken Saudi phone numbers transcribe correctly only a third of the time, so a lead is confirmed by read-back or a typed field rather than trusted. The faster speech model won on real reply lengths by about 830 ms per turn, so it is the one in use. Greetings are pre-rendered, so an idle visitor costs nothing. Four guards close the open-microphone tap: thirty seconds of silence, a switched tab, twenty-five turns, and daily caps. Sixty-five tests, each one written from something that actually broke.

    04What shipped

    1. 01
      Answers only from a curated knowledge base
    2. 02
      Voice with barge-in and zero-cost pre-rendered greetings
    3. 03
      Lead capture with a phone-number read-back gate
    4. 04
      Server-owned consent and daily cost caps
    • AI agents
    • Python
    • Django
    • API engineering
    • Security engineering
    SUAD AISuad AI admin: provider settings, no vendor hard-wired
    Suad AI admin: provider settings, no vendor hard-wired Suad AI admin: system status checks
    Role
    An AI employee for support, appointments, orders and FAQs, with human hand-off
    Languages
    Follows the chosen models — English and Arabic with local or cloud providers
    Channels
    Chat, voice, phone calls (SIP and Twilio), admin console

    02 / AI assistantsActive

    Suad AI

    Own product

    01The system

    A local-first AI employee a business can run on its own hardware.

    Suad chats, listens and speaks, answers phone calls, reads a knowledge base, uses typed business tools and follows task playbooks — support tickets, appointments, order lookups, FAQs, hand-off to a human. Every provider is swappable: language models, speech in and out, telephony and memory, with nothing hard-wired to a vendor.

    02What I did

    Architected and built the platform: the FastAPI backend and provider adapters, the agent runtime, a permission engine, YAML task recipes, telephony adapters for SIP and Twilio, and the React admin with its setup wizard, playground, voice test and call logs.

    03How I built it

    Local-first by default so it runs entirely offline with Ollama, Whisper and Piper. The model only ever emits tool requests; typed handlers run the queries, so it can never write SQL. Every tool call passes a permission engine — roles map to scopes, each action carries a risk level of safe, confirm, approval or blocked — and every decision lands in an immutable audit log. The test suite runs green with no cloud account at all.

    04What shipped

    1. 01
      Swappable LLM, speech, telephony and memory providers
    2. 02
      Permission engine with risk levels and an audit log
    3. 03
      YAML task playbooks for everyday business work
    4. 04
      Runs fully offline with local models
    • AI agents
    • Workflow automation
    • Python
    • API engineering
    • JavaScript
    • Security engineering
    NZUM SERVICES ERPSYSTEM PORTRAIT
    NZUM Services ERP

    An operations platform for a services company: HR, ticketing, CRM and AI media in one place.

    Private system — no public screens

    03 / PlatformsActive

    NZUM Services ERP

    Built end to end

    01The system

    An operations platform for a services company: HR, ticketing, CRM and AI media in one place.

    Recruitment from public job posting to signed employee, attendance and document-expiry tracking, ticket lifecycles with SLAs and escalation, a sales pipeline, team chat, and a management dashboard — with AI and machine-learning media tools folded into the same system.

    02What I did

    Built the platform and its modules, and kept it in production use.

    03What shipped

    1. 01
      HR: hiring, attendance, document expiry
    2. 02
      Ticketing with SLAs and escalation
    3. 03
      CRM and sales pipeline
    4. 04
      AI and ML media tooling
    • Workflow automation
    • Python
    • Django
    • Data pipelines

    05 / More systems

    Also built.

    Maria Technolahi on the Altria Call Center site

    AI assistantsLive

    Maria Technolahi

    A multilingual AI customer-service agent for Altria Call Center.

    Built & evolved. Built the AI agent and continued improving its product experience and capabilities.

    Live (opens in a new tab)
    Executive Systems home with the AI assistant entry point

    AI assistantsLive

    Executive Systems

    A bilingual IT-services platform with a consent-aware AI assistant.

    Built end to end. Built the website and its AI-assisted enquiry experience, including text, voice and human handoff.

    Live (opens in a new tab)
    Chart Idea home

    PlatformsLive

    Chart Idea

    An e-learning and financial platform with Sharia-compliant stock screening.

    Built end to end. Built the platform: course management, the screening engine, billing, and video delivery.

    Live (opens in a new tab)Source
    Locum Meds home with the job search journey

    PlatformsLive

    Locum Meds

    A clearer digital recruitment journey for UK healthcare staffing.

    Rebuilt & improved. Rebuilt and improved the website experience, making the core candidate and employer journeys easier to discover and use.

    Live (opens in a new tab)
    CageFit Gyms home

    PlatformsLive

    CageFit Gyms

    A complete web and membership journey for a 24/7 combat-sports gym.

    Built end to end. Designed and built the website end to end, from service discovery to membership action.

    Live (opens in a new tab)

    06 / Career

    Places my work has shipped.

    Above: the organisations the work shipped with, each carrying what I did there. Below: the tools it was built with.

    Coinova logo CoinovaWebsite build
    Locum Meds logo Locum MedsRebuilt & improved CageFit Gyms logo CageFit GymsBuilt end to end
    ChartIdea logo ChartIdeaWebsite build
    ATCL logo ATCLWebsite build
    Python PythonBackend systems & automation
    Django DjangoProduction web platforms
    FastAPI FastAPIService APIs
    PostgreSQL PostgreSQLPrimary datastore
    Redis RedisCache, queues & sessions
    Celery CeleryBackground work
    JavaScript JavaScriptProduct interfaces
    TypeScript TypeScriptTyped front ends
    React ReactAdmin & app interfaces
    Tailwind CSS Tailwind CSSDesign systems
    HTML HTMLSemantic structure
    CSS CSSInterface craft
    Node.js Node.jsTooling & build
    OpenAI OpenAIAssistants, speech & vision
    Ollama OllamaSelf-hosted models
    PyTorch PyTorchMachine learning
    NumPy NumPyNumerical work
    pandas pandasData pipelines
    Qdrant QdrantVector search & RAG
    LiveKit LiveKitRealtime voice
    Twilio TwilioTelephony
    Docker DockerReproducible environments
    NGINX NGINXEdge & reverse proxy
    Linux LinuxServers & operations
    Ubuntu UbuntuProduction hosts
    Bash BashOperations & deploys
    Git GitVersion control
    GitHub GitHubOpen source
    Cloudflare CloudflareDNS & edge
    Let's Encrypt Let's EncryptTLS everywhere
    Playwright PlaywrightBrowser automation & tests
    SQLite SQLiteLocal & embedded data
    NVIDIA NVIDIACUDA, NVLink & MIG
    Kubernetes KubernetesContainer orchestration
    vLLM vLLMHigh-throughput model serving
    Hugging Face Hugging FaceTransformers, PEFT & Diffusers
    Jupyter JupyterGPU-backed notebooks
    Prometheus PrometheusMetrics & alerting
    Keycloak KeycloakIdentity, OIDC & MFA
    WireGuard WireGuardPrivate networking
    Traefik TraefikAuthenticated reverse proxy

    07 / Capabilities

    What I can build and run.

    Not everything here appears in the work above — much of it is private. This is the range I am deployment-ready on, by domain.

    01

    Core engineering

    5
    • PythonBackend systems, services and automation
    • DjangoProduction web platforms
    • FastAPIService APIs and agent runtimes
    • API engineeringReliable service integrations
    • Bash & shell automationOperations, deploys and glue
    02

    Interfaces

    3
    • JavaScriptInteractive product experiences
    • ReactAdmin consoles and app interfaces
    • Responsive UXAccessible interfaces across devices
    03

    AI & machine learning

    6
    • AI agentsTask-focused conversational systems with tools
    • RAG pipelinesRetrieval-grounded answering over real documents
    • Voice interfacesSpeech in and out, barge-in, telephony
    • PyTorchTraining and inference
    • Hugging Face stackTransformers, PEFT, Diffusers, DeepSpeed
    • Workflow automationOperational tooling and integrations
    04

    GPU & AI infrastructure

    7
    • NVIDIA GPU stackDatacenter drivers, CUDA, cuDNN, Container Toolkit
    • NVLink & MIGFabric validation and GPU partitioning for multi-tenancy
    • vLLM servingOpenAI-compatible endpoints, KV cache, tensor parallelism
    • Embeddings inferenceEmbedding generation at scale
    • JupyterHubGPU-backed notebooks with per-user limits
    • Model catalogueWarm-set residency and on-demand loading
    • LLM gateway & meteringProxy routing with usage accounting and quotas
    05

    Infrastructure & orchestration

    6
    • LinuxHardening, kernel tuning, service minimisation
    • SlurmWorkload management, GPU accounting, fair-share scheduling
    • KubernetesContainer orchestration
    • DockerReproducible environments and compose stacks
    • Storage & LVMVolumes, mounts, permissions and quotas
    • nginx & TraefikAuthenticated reverse proxies and TLS
    06

    Observability & operations

    4
    • PrometheusMetrics collection and alerting
    • GPU telemetryDCGM exporter and utilisation accounting
    • Centralised loggingLoki with retention policies
    • Runbooks & handoverOperational procedures, user guides, KT
    07

    Security & identity

    3
    • KeycloakOIDC, MFA enforcement, RBAC tiers
    • WireGuard & OpenVPNPrivate, authenticated remote access
    • Security engineeringSecure-by-default product design
    08

    Data systems

    4
    • PostgreSQLPrimary datastore
    • Redis & CeleryCache, queues and background work
    • Vector searchQdrant-backed retrieval
    • Data pipelinesCollection, transformation and delivery

    38 capabilities across 8 domains · the tools below are the ones with a mark worth showing

    08 / Field notes

    Systems, not just screens.

    I've spent more than 14 years working across programming, software development and systems that solve practical problems. I build the product, improve the operation behind it, and share what I learn through 7AZZANI.

    — Abdullah AzzaniFounder of 7AZZANI

    1. 01

      Start with the operation.

      Understand the real workflow before deciding what the interface or the assistant should do.

    2. 02

      Measure before deciding.

      Observe the failure, the latency or the hand-off before making a design choice.

    3. 03

      Build for the next change.

      Maintainable foundations, honest limits, and room to evolve after launch.

    See the full capability matrix ↑