About
Al Marfa is an engineering studio that builds software with AI agents and publishes how it is checked.
We design the harnesses, review panels, and evaluation methods that let agents do real work, run them on our own products, and write down what happened. Frontend architecture is where we came from and still part of what we do.
Why This Matters
The problem
Most teams using AI agents have no way to tell good output from confident output. They either review everything by hand, which is slower than not using agents, or trust runs they cannot check.
Our approach
We put the effort into the structure around the model: what it may touch, who checks it, and what proof it returns. Then we review the disagreements, sample the rest, and keep the record.
Core Principles
What we hold to when agents do the work.
Evidence before claims. A method is offered only after it has run on our own products at full scale, and the numbers are published.
Independence makes a review. A second opinion counts when it comes from a different context, and often a different vendor.
Agents do the volume, people decide what ships. Human time goes to disagreements, samples, and the release, and the decision is recorded.
Some decisions are never delegated. Where a wrong answer carries real weight, the rule is written down and the pipeline enforces it.
How We Operate
A studio that publishes
Al Marfa works in the open like a technical publication. Case studies keep their numbers, including the misses, because that is what makes the next decision better. The same discipline applies to the products we operate.
Credibility through the record
Our credibility is the record: products in production, pipelines with their verdicts pinned, and tests that fail when something is unverified.