Applied AI Engineer
I build and evaluate the
knowledge layer behind AI
Years of Tier 3 support escalations taught me to tell a plausible answer from a correct one. I do that for AI systems now: retrieval grounding, eval harnesses, and the curated knowledge an assistant actually answers from.
Now
What I'm working on
I work on a Claude-powered advisor for membership-site operators. My side of it is answer quality end to end: the curated knowledge the advisor draws on, the retrieval that grounds it in a customer's real data, and the shipped fixes when an answer is confident and wrong. That runs from writing and reviewing the corpus to the product code that decides how a citation renders and what the advisor is allowed to assert when the data is missing.
Outside that, I run my own AI setup as a working system rather than a chat window: Claude and Cursor daily, a file-based memory layer that survives between sessions, and self-improving agent loops on a schedule. Most of what I publish comes out of things that broke in that setup.
Updated August 2026
How I work
Getting an assistant to say only what the data supports
Years of escalations at Tier 3 trained one instinct: telling a plausible answer apart from a correct one. Applied to an AI product, that turns into two jobs.
Answers tied to the customer's own data
Every answer the advisor gives has to rest on the customer's own data and name where it came from. My side is keeping that citation path honest, and stopping the model asserting anything the data doesn't actually support.
Caught before the user sees it
Grounding evals that fail when a prompt or retrieval change lets a bad answer through, so fixing one answer can't quietly break six others.
What I do
Where I focus
The work sits where an AI system meets the knowledge it depends on, and where somebody has to prove the answers hold up.
Evaluation and testing
Knowing whether the answer is actually right
- Eval harness design and regression coverage for prompts and retrieval
- Scoring across multiple runs so run-to-run variance doesn't read as signal
- Reading results honestly, including the ones saying a change didn't work
Retrieval and knowledge curation
Deciding what an assistant is allowed to know
- Curating and reviewing the corpus an assistant answers from
- Grounding answers in sources, with citations that survive a check
- Review processes that keep weak or contradictory content out
Agents and context engineering
Making sessions, memory, and tools behave
- Context and memory design so agents keep state across sessions
- MCP tooling and agent workflows
- Permission boundaries that act as controls, not polite requests
Working on something similar? Send me a note .
Stack
What I work with
Claude and Cursor daily, with self-improving agent loops that run on a schedule rather than waiting for me to prompt them.
Work
What I've built
That work ships into a live product in a private repo, so there is nothing to click on those — the write-ups say what the problem was and what changed. Everything below them is public and open source.
Finding that "cited" wasn't measuring citation
The Problem
The product measured which knowledge-base entries were performing by counting citations. Tracing the retrieval path while building usage analytics, it turned out nothing ever parsed the answer text for citation markers. "Cited" had silently meant "was in the top-k handed to the model" — never "the model actually used it." Every downstream decision about which content was working rested on that.
The Solution
Built a retrieval-event log that records the full set of entries retrieval considered for each answer, not only the ones returned, so the difference between retrieved, returned, and genuinely used becomes visible. Fed that into a nightly per-entry rollup covering citation counts, last-cited, feedback-bearing citations, and helpful rate across all-time and 30-day windows.
The Outcome
Content review runs on a signal that means what it says. The measurement bug would have kept quietly misreporting which knowledge was earning its place.
Citation integrity in advisor answers
The Problem
The advisor was emitting literal bracket tokens inline in its answers, always straight after a claim about live customer data. They looked like citations but pointed at nothing, so they rendered as raw text on exactly the grounded answers the product exists to give.
The Solution
The system prompt told the model to cite using bracket notation matching the provided context. Scoped that broadly, the model improvised brackets out of the context section labels, and the renderer only stripped numbered ones. Narrowed the instruction, replaced the renderer with a dedicated strip layer, and added a grounding eval that fails if a descriptive token reaches an answer.
The Outcome
Real citations render, invented ones never reach the user, and the eval keeps it from regressing quietly.
Refusing to assert a false zero
The Problem
Some billing connectors report members but no transactions, because the source plugin keeps no money ledger. "We have no revenue data" and "revenue is zero" both came out as $0.00, so the advisor would state a zero it had no basis for. That is the most damaging class of AI error: confident, specific, and wrong in the direction an operator acts on.
The Solution
Declared which platforms actually carry a billing ledger, added a revenue-coverage layer running through the profile and rollup paths, and had the response layer mark coverage as unavailable with an explicit instruction not to report, infer, or estimate revenue for those sites. Solved for the whole class of connector rather than one platform at a time.
The Outcome
The advisor states that the data isn't there instead of printing a number that looks like an answer.
File-based agent memory
The Problem
LLM sessions start cold every time, so every morning begins by re-explaining the same context before any work happens.
The Solution
A file-based memory layer with a defined structure for facts, projects, and preferences, plus rules for what gets written and what stays out.
The Outcome
An agent picks up the thread from the previous session instead of being re-briefed. Running my daily work for months, and now published as an open-source template.
Permission boundaries for agents
The Problem
Safety rules written into prompts read like controls but behave like suggestions, and the difference only shows up when one gets ignored.
The Solution
Audited every safety rule in my own AI setup, sorted prose rules from enforced ones, and mapped what it takes to turn a request into an actual boundary.
The Outcome
A working model of where instruction ends and enforcement begins, written up with the autonomy ladder it produced.
Administrator Rescue Link — Members plugin
The Problem
Remove one critical capability from the Administrator role and you lock yourself out of WP Admin completely. Recovery meant direct database access through phpMyAdmin or WP-CLI, which most site owners can't do safely.
The Solution
Contributed the Administrator Rescue Link to the Members plugin (300,000+ active installs). It emails a time-limited magic link to the registered admin address; clicking it restores Administrator capabilities through a dedicated front-end endpoint. Token verification, per-user rate limiting, Superadmin support for multisite, and a refactor of the rescue flow.
The Outcome
Shipped in the current Members release. Locked-out administrators recover in under two minutes without touching the database or opening a ticket.
More WordPress work — bulk PDF invoicing, staging email guardrails, per-member content rules, and forward-only access — lives in public repos on GitHub.
Browse repositories on GitHubProof
Verified, not asserted
Five-star Trustpilot reviews
Public reviews that name me, across MemberPress and Leap13. Every one is searchable on Trustpilot.
View profilePlugins on WordPress.org
Published under my own account, alongside code contributions to the Members plugin.
View profileRemote across four regions
Delivered across time zones with teams in the USA, UAE, Egypt, and Europe.
Reviews
What customers wrote
Public Trustpilot reviews that name me, quoted verbatim. Search “Omar” on the MemberPress page to read all of them.
“Working with customer support in MemberPress, and Omar in particular, has been shockingly great. No quick, confusing answers that leave you with more questions — long, detailed explanations that completely clarify the situation. And when I had an unusual situation, they developed a plug-in for me to cover it! Absolutely outstanding!”
“My latest experience when I needed guidance on how to do something I was quite willing to upgrade to Pro but Omar explained how to carry out what I needed to do without upgrading. Now that's unheard of as most companies would gladly just point you towards the upgrade option! 10 out of 10 Guys, keep up the great work.”
“The support replies quickly and gives specific, detailed and practical solutions. As a software developer I'm using different tools and having a good support is a key point for me to use a software long term over multiple years.”
“I've designed and managed several WordPress membership site projects with MemberPress and have always found their support to be top quality. With a recent issue I was facing, Omar went the extra mile to find a solution and resolve the problem.”
“I spoke with Omar who gave fast and correct support to solve my client's issue immediately. Really impressive. Thank you Omar!”
18 public Trustpilot reviews name me across MemberPress and Leap13.
Background
Where I've worked
Support engineering, then consulting, now the knowledge layer of an AI product.
Download CVAI Knowledge Engineer
Caseproof (MemberPress)
Own knowledge quality for an AI advisor product: what the assistant is allowed to know, how the curated corpus it answers from gets written and reviewed, and how anyone tells whether its answers hold up.
- • Author and review the curated corpus behind the assistant
- • Run the review process that keeps weak content out of it
- • Work on grounding and answer quality with the AI engineering team
Tier 3 Support Engineer
Caseproof (MemberPress)
Handled the escalations that got past Tier 1 and Tier 2: subscription logic, payment gateway integrations, and platform-critical bugs on a global SaaS product.
- • Tier 3 escalations on the tickets nobody else could solve
- • Stripe and PayPal subscription and webhook debugging
- • 18 public five-star Trustpilot reviews naming me
WordPress Support Engineer
Caseproof (MemberPress)
Supported WordPress-based SaaS implementations at scale, working directly with customers on membership and payment problems.
- • Troubleshooting across large WordPress and MemberPress installs
- • Complex membership, subscription, and payment resolution
- • Built the ticket volume that the Tier 3 work grew out of
WordPress Technical Consultant / Lead Developer
SwiftX (Self-employed)
Technical consultant and lead developer on WordPress-based platforms and client solutions, alongside the day job.
- • Turned business requirements into technical specifications
- • Architected and delivered client platform work
- • Led delivery across client engagements
WordPress Developer
Leap13 / LeapWorx
A hybrid role across WordPress development, technical support, and technical content for client-facing platforms. Earned three verified five-star Trustpilot reviews for client work.
- • Custom PHP and JavaScript plugin and theme development
- • WordPress and WooCommerce client platforms
- • Where the support side of the work started
Contact
Open to applied AI
engineering work
Evals, retrieval, agent tooling, knowledge systems. Remote, based in Cairo. Also happy to talk about anything in the writing.