Muhamed Alarbi

LLM systems

1. LINGUA

   LLM-driven, intent-aware scheduling for the cloud-edge continuum

   Operators describe what they want in plain language, such as lower latency or lower cost, and a large language model produces the deployment plan. A guardrail checks every plan and sends violations back to the model until the plan is valid, so changing the objective means editing a prompt, not rewriting a scheduler.

   Across 20 models from 7 providers, the strongest came within about 1% of the optimum found by a MILP solver.

   Paper to be submittedWith AWS Edge Computing

2. MAPeR²

   Memory-augmented LLM agents for personalized energy recommendations

   A multi-turn LLM agent that gives households day-ahead recommendations for when to run their appliances, taking time-of-use electricity prices into account. It keeps track of the conversation, recalls past interactions from a vector database, and turns each finished session into lessons about what worked and what to avoid.

   With episodic recall, recurring schedules were accepted in a single turn with no corrections.

   IEEE WF-IoT 2026 · accepted

3. Archive Assist (دليل)

   Arabic document intelligence · Digital Assets

   A platform I design and build end to end at Digital Assets. Scanned Arabic official documents and spreadsheets become a searchable archive that answers questions with citations to the exact page or row, with briefing notes and decision memos built on top. Sensitive documents are processed only by on-premise models.

   Go, PostgreSQL, Meilisearch and SolidJS, with a model layer that works across seven LLM providers.

   Digital Assets · 2026–present

4. Marketplace search

   Semantic and hybrid product search · Digital Assets

   Product search for an Arabic e-commerce marketplace that combines keyword matching with OpenAI embeddings in Meilisearch. An LLM generates search keywords for every product, covering Modern Standard Arabic synonyms, Libyan and other dialect words, and English, so shoppers find products in the words they actually use.

   A relevance cutoff set from measured score bands, related-product recommendations from embedding similarity, and a database fallback if the index is down.

   Digital Assets · 2026

5. Content moderation

   AI moderation for an online marketplace · Digital Assets

   Every product listing, review and seller profile is checked against the marketplace’s policy by an LLM with strict structured output, from illegal goods and scams to off-platform contact details and political content. It reads Libyan dialect and Arabizi, and also assigns each product’s category and shipping size.

   Runs asynchronously through a transactional outbox and job queues; only high-confidence verdicts are acted on automatically, and the rest go to human review.

   Digital Assets · 2026