About

policyon

AI governance, from policy to production

policyon is read Policy. On. - policy switched on, enforcement live, not a document in a drawer. It is the working name for one practice, and twenty years in software sit behind it. We lead AI delivery: ambiguous problems taken through discovery to systems that run in production. We front client engagements from the first conversation to an approved proposal, then own the roadmap for what gets built.

The through-line across all of it is a suspicion of unchecked claims, including our own. That is why we build enforcement rather than documentation, why the accuracy standards we set are designed to be falsifiable, and why this site publishes its own corrections.

01 · What we do
Focus
AI delivery, applied AI governance, agentic systems and evaluation.
Engagements
Individuals, teams and organisations, from a single system to a whole estate.
Working
Remote from Egypt, hours overlapping UK, EU and US teams.
Languages
English, professional. Arabic, native.
Currently working on
  • An AI platform that reads engineering drawings and extracts bills of materials, deployed and live.
  • A governed agent harness, armed on real projects rather than demos.
  • Responsible-AI deliverables mapped to recognised frameworks, with the data handling settled before any model sees a document.
  • An AI adoption remit across product teams: where AI belongs, the guardrails on output, and the onboarding that makes it stick.
02 · The author

Twenty years, mostly on systems where being wrong was expensive.

policyon is run by Ismail Saleh, a senior AI product lead in the industry since 2003. The current work is client-facing lead on AI engagements from discovery through to approved proposal, roadmap ownership for what gets built, and an AI adoption remit across product teams.

Before that, six years of enterprise full-stack work on intelligent document capture and routing platforms, and a decade before that on loyalty systems, commerce and the desktop software nobody writes about - installing it, then answering the phone afterwards. That is where the habit came from: when you are the person who takes the call, you stop shipping things you have not checked.

The role-by-role record is on LinkedIn. What the practice has actually built is on the case studies page, with its receipts.

History is only worth telling if it says something about the next engagement. These are the two things it says: what we can pick up, and what we are difficult about.

03 · What working with us is like
Technical

.NET and C#, ASP.NET Core and MVC, Python 3.11+, TypeScript and JavaScript, Angular 2+, SQL Server and Postgres, REST APIs and SignalR, Docker, AWS. SOLID, TDD and design patterns, Git, Jira, Agile and Kanban.

AI stack: LangChain and LCEL, Claude, OpenAI and Gemini, ChromaDB, OpenAI and HuggingFace embeddings, Firebase Firestore, Firecrawl, MCP servers.

Agentic work: multi-agent orchestration, ReAct agent design, RAG pipeline architecture, system prompt design, prompt evaluation and testing, human-in-the-loop design, guardrails and verification gates, agent harnesses, retrieval, evaluation harnesses, workflow automation.

Day to day: Cursor, Antigravity, VS Code with Claude, LangChain Hub, Pyright, Poetry.

How we prefer to work
  • Give us the ambiguous problem, not the pre-written ticket.
  • Let us define what correct means before the build starts.
  • Expect the gap in our own work to be in the report, near the top.
  • Decisions get written down the day they are made, or they did not happen.
Next

Open to conversations.

Delivery leadership, AI governance and assurance, agentic systems, or a walkthrough of the harness against your own repository.

If you would rather read the engineering before speaking to anyone, that is the right instinct. The case studies carry the measurement and the failure reports; the code behind the current harness is available on request.

04 · How this site was made

Most of the writing here was produced by AI systems working under our direction. Including this page.

Written and directed by Ismail Saleh, who is accountable for every claim on this site.

That is the honest description of how this practice works, so hiding it would fail the standard the rest of the site argues for. It also decides what you should take from the site, which is why it comes with a grading rather than an admission on its own.

Measured

trust this

Counts, dates and file states. Checked against the artefacts rather than remembered, and re-run whenever a first figure looked generous rather than accepted because it flattered.

Ours, plainly

trust this

The ten questions on the overview, typed verbatim in live sessions and traced back to the turn. The protocols they specify: a controlled before-and-after benchmark, a named regression baseline that has to pass before anything is touched, a second-opinion gate enforced by a pre-write hook rather than requested politely, verification of a tool’s self-report against repository state. The judgements - what to kill, adopt, re-scope, refuse. The corrections.

Commissioned

take with the grading

Much of the framework prose - the rubrics, the catalogues, the taxonomies - was drafted by models answering our questions. Where a standard is ours, it was set in our own words first and the wording came after. Where it is not, we set the problem, refused the first answers, and decided what shipped. We did not write every sentence. Saying so is the same standard we would hold a client’s disclosure to.

Withheld

take with the grading

Anything naming a client, an employer or a specific system. Also any figure tied to a repository whose ownership is not yet settled in writing.

Why this is the section that matters

A model will produce a governance framework for anyone who asks. What it will not do is tell you the framework is wrong, or notice that the gate depends on the module whose failure triggers it. It will not refuse a number just because nobody can source it.

That part is the job, and it is the part of this site that is not commissioned.