AI automation engineering · Commercial systems

Less busywork.
More business.
Built to connect.

I'm Abdullah. I bring commercial experience and hands-on automation together to connect the work between your inbox, CRM and operations.

Based in Pakistan · Open to remote opportunities

THE CONNECTED OPERATIONINTERACTIVE DEMOS BELOW
↳
A conversation comes inEmail, call transcript or customer request
✳
AI turns context into structureIntent, fields, priority and next action
✓
The right workflow takes overCRM update, draft reply or approval queue
Human judgment where it matters.01 → 02 → 03
Business context
behind the examples
marlead.feliz pakistanQORDEN AI
01 / Project examples

Real business context.
Practical system thinking.

Three business scenarios that connect my commercial background with the workflows I focus on building.

Try the ComplyOne business verification workflow ↗
Marlead / Revenue operations
Lead signal
→
Qualify
→
Next action

A pipeline that keeps moving.

Turn scattered prospect information into structured CRM records and a clear follow-up queue.

Feliz Pakistan / Customer operations
Customer email
→
Classify
→
Draft + route

Give every request a next step.

Separate routine product and delivery questions from issues that need a person to step in.

Qorden AI / Commercial intelligence
Call transcript
→
Extract
→
CRM draft

Make conversations actionable.

Capture buyer requirements, commercial signals and next steps without rebuilding call notes by hand.

These are illustrative system designs grounded in my business experience, not claims of production deployments at these companies. Demos use synthetic data. Client-specific implementation details and outcomes are available only where permitted and verified.

02 / The workflow lab

Don't just read it.
Run through it.

Explore the inputs, decisions and outputs of four automation patterns. Browser-based simulations, no accounts or live integrations required.

Run the Dubbix localization demo ↗
Choose an input and run the workflow.

Deterministic demonstration logic, not an LLM API. Nothing is sent, scheduled or written to a CRM. Avoid entering confidential information.

One video.
Three release decisions.

Run a simulated dubbing job for a product launch. Repair timing, resolve voice consent or replay a failed render. Each language clears its own release gate.

01 · Ingest02 · Translate03 · Quality gates04 · Render05 · Package
Event trace

Portfolio prototype using sample translations, simulated audio timings and mock renders. Caption and manifest exports are generated locally.

Business onboarding · Workflow demo

A business signs up.
Who needs to know?

When a company submits its registration details, ComplyOne starts a Sumsub business check automatically. If the workspace requires a second check, it also routes the case to VOVE ID. A possible screening match puts onboarding on hold and gives the workspace owner one case to review.

New business application

Northstar Studio Ltd
United Kingdom · Demo registration NS-042

Workspace owner inbox

Submit the sample application to see how the case is routed.
Provider event log

    This portfolio prototype uses fictional company data and simulated API responses and alerts. A live integration would call providers from the backend, authenticate provider callbacks and restrict case evidence to authorized reviewers. Sumsub API · VOVE ID KYB

    03 / Engineering notebook

    Beyond the happy path.

    Executable browser examples and implementation blueprints: explicit contracts, controlled side effects and observable failures.

    01 / IngestionSigned webhook, event ID, tenant context
    02 / ProcessingNormalize, redact, extract structured fields
    03 / ValidationSchema checks, source evidence, policy rules
    04 / ReviewApprove a versioned proposed action
    05 / ExecutionScoped CRM write, idempotency, reconciliation
    06 / ObservabilityCorrelation ID, latency, cost, retry and audit log
    Executable example / delivery control

    When the integration fails.

    Run a small event processor with a duplicate event, an API failure or an unapproved action. The state machine below runs locally in JavaScript.

    Awaiting event...

    Retries are simulated immediately. Production requires durable storage, a queue and atomic uniqueness constraints.

    Qorden AI / extraction contract

    Facts with evidence.

    A CRM extraction should distinguish missing information from facts, with a trace back to the source.

    {
      "schema_version": "1.0",
      "event_id": "call_demo_001",
      "company_id": "demo_company",
      "requirements": [{
        "value": "Multilingual speech analytics",
        "evidence": "We need multilingual speech analytics",
        "source": "transcript:turn_1"
      }],
      "budget": null,
      "timeline": null,
      "next_action": "schedule_technical_review",
      "write_status": "pending_approval"
    }

    Validate types and allowed values before proposing a write. A model-generated confidence score alone should never authorize a CRM update.

    Feliz Pakistan / retrieval design

    Answers grounded in approved data.

    Separate product and policy retrieval from order access. A product question can use approved public content. Order details require a verified customer and order match.

    • Attach source ID, content version and update timestamp to retrieved passages.
    • Filter by business and document permissions before retrieval.
    • Return a clarification request when evidence is missing or conflicting.
    • Escalate refunds, payment disputes and safety questions.
    request → classify → permission check
    product route: approved product index
    order route: verified order lookup
    sources → grounded draft → policy check → review

    Proposed architecture. No vector database or private order data is connected to this portfolio.

    Marlead / evaluation harness

    Test the decisions, not the wording.

    Run a deterministic policy regression suite. These checks exercise routing rules; they do not measure an LLM's quality.

    Ready to evaluate: opt-out, closed deals,
    missing identity, missing approval,
    valid next action and invalid timing.

    For a real model pipeline: add labeled fixtures, field-level precision and recall, unsupported-claim checks, prompt-injection cases, latency and cost budgets.

    These engineering examples were built for this portfolio. Architecture proposals and sample schemas are not production deployment claims.

    04 / Open-source references

    Building blocks worth knowing.

    External GitHub projects relevant to the designs above. Linked as references, with credit to their maintainers.

    Orchestration

    n8n

    A fair-code workflow automation platform. Relevant to webhook ingestion, API integrations and review handoffs in the Marlead example.

    n8n-io/n8n ↗
    Evaluation

    Promptfoo

    Tools for evaluating and red-teaming LLM applications. Relevant to regression tests for extraction and grounded customer replies.

    promptfoo/promptfoo ↗
    Data privacy

    Presidio

    PII detection and anonymization tooling. A reference for reducing personal data in messages and transcripts before model processing.

    data-privacy-stack/presidio ↗

    Third-party repositories. No authorship, contribution, affiliation or prior implementation is implied. Review each project's license and deployment requirements before use.

    03 / How I think

    The workflow comes first.
    The tools follow.

    Map the business decision.

    Define the trigger, source of truth, owner and what a successful outcome looks like.

    Connect structured actions.

    Use APIs, validated fields and explicit routing. Give ambiguous requests a human review path.

    Plan for the messy parts.

    Design for retries, duplicate events, missing data, approval boundaries and a useful audit trail.

    Commercial perspective

    I understand why the system needs to exist.

    My experience spans commercial leadership at Qorden AI, GTM advisory through Marlead and running Feliz Pakistan. That means seeing the work from the sales team, the operator and the customer's side.

    I focus on the handoffs: where a conversation becomes a task, a task becomes a follow-up, and a follow-up either moves the business forward or gets lost.

    n8n / MakeLLM APIsHubSpotGmail + CalendarShopifyWebhooks + JSON

    Architecture options shown for these demos. Specific tool experience and production references can be discussed directly.

    04 / Next conversation

    Where does your team
    lose the most time?

    Let's start with one workflow and a clear outcome.

    Email Abdullah ↗

    abdullahismail100@gmail.com