01

What I build

  • Model-driven search

    The model calls licensed data tools through MCP, under hard rules that cap how many credits a run may spend.

  • Ranking and filters

    Local ranking with strictness levels and progressive retries when a strict search finds nothing.

  • Human approval

    Results wait for a person to approve them before anything is written to the CRM.

  • Long-running jobs

    Resumable bulk jobs, async webhooks, rate-limit backoff and recovery pages for anything that fails.

  • Cost visibility

    Logs and stats per run, so you know what the AI did and what it cost.

02

How I work

  1. Step 1
    Pick one job

    AI where it saves real time: search, classification, enrichment. Plain code for everything else.

  2. Step 2
    Set hard limits

    Credit and rate rules in the prompt and in the code, so a run cannot overspend.

  3. Step 3
    Keep a human in the loop

    Approval before data reaches a customer-facing system.

  4. Step 4
    Make it recoverable

    Every job resumable, every failure logged with enough detail to retry.

03

Selected work

Client names are redacted to protect their systems. The work and the numbers are real.

  • Built · Iterating

    B2B lead intelligence pipeline

    Events and exhibitor companies come in from a HubSpot custom object or a CSV. The OpenAI Responses API drives a licensed B2B data provider through MCP to find companies and people, under hard credit rules. Results are ranked locally, emails are found through parallel async requests with webhooks and backoff, and approved records are pushed to HubSpot as associated companies and contacts.

    So far: about 296,000 records processed across 86 runs, and more than 37,000 approved contacts pushed to HubSpot. Behind it: resumable jobs, diagnosis and recovery pages, and data lifecycle tools for archiving and restoring.

    • OpenAI Responses API
    • MCP
    • PHP
    • MySQL
    • HubSpot API

    Read the case study →

  • Teaching

    AI lessons in Arabic

    Short AI series on my YouTube channel, next to the long-running programming courses.

04

Quick answers

Will AI write directly into my CRM?

Not without a rule you choose. In my pipelines a person approves results before they are pushed.

How do you keep AI costs under control?

Hard credit limits per run, local ranking so the model is called less, and logs that show what each run used.

Do you build chatbots?

My focus is AI inside business processes and data pipelines. A chatbot is worth it only if it has a reliable source of knowledge behind it.

05

Related