Research direction of Engineer Mohanad Osama Zahran

A programme centred on decisions

Mohanad Osama Zahran leads the concept and research direction of JARVIS and System Brain at ZEISMO Research & Lab. The programme asks how a software agent can take responsibility for a defined task while remaining constrained by evidence, permissions and a record of its actions. This makes the unit of study a decision within a workflow, rather than a broad claim about intelligence.

Three connected research questions

First, how should a role-specific agent represent its responsibilities and available skills? Second, how should a coordinating system use memory without turning old assumptions into current facts? Third, when should an agent act, request help or stop? These questions connect the architecture in the JARVIS report to the lab’s follow-on decision study.

JARVIS and System Brain have different roles

JARVIS is proposed as an exclusively English-speaking interface. System Brain is the proposed coordination layer for specialised software agents. A role cell represents a software responsibility, with instructions, tools and memory. The research uses this separation to examine what an interface can propose and what the underlying system is authorised to execute.

The published decision pilot

When Should an AI Employee Decide, Ask for Help, or Stop? uses 48 synthetic scenarios, divided into 12 development cases and 36 evaluation cases. The computational pilot compares fixed rules, one model call, a three-call System Brain pattern and three-call self-review. Published outputs and calculations can be inspected; the original human explanation-review endpoint remains unperformed.

From ambition to a measurable question

The programme’s specialist-coordination ambition needs a defined task and a fair comparison. Accuracy, unsafe actions, escalation quality, latency and cost measure different things. A claim about many specialists or a general human-intelligence multiplier cannot be inferred from software tests. The research direction is to make those comparisons specific enough to test.

How to follow this work

Start with the technical preprint for architecture and recorded experiments, then read the decision-dataset brief for the follow-on method. New results should identify the exact version, evaluation conditions and evidence package. Research enquiries can address a specific method, scenario or replication question through the lab’s contact page.