Luay Gharzeddine

Senior AI engineer building production LLM systems: multi-agent orchestration, text-to-SQL, retrieval, and the evals and guardrails that make them safe to ship.

Beirut, Lebanon UTC+2

Portrait of Luay Gharzeddine

I build LLM systems for regulated industries. I designed the orchestration layer for a bank's customer-facing AI assistant, covering 20+ banking workflows; built clinical documentation tools that cut physicians' report drafting from 20 minutes to under 5; and built a multi-agent automation platform where I reduced LLM costs 8x. My research on multi-agent task decomposition was published in Transactions on Machine Learning Research (2026).

I care about measuring things properly. Every project below ships with reproducible evaluations, honest baselines, and the numbers to back its claims.

Research

TMLR 2026

Complete Cyclic Subtask Graphs for Tool-Using LLM Agents: Flexibility, Cost, and Bottlenecks in Long-Horizon Workflows

Luay Gharzeddine, Samer Saab Jr. Transactions on Machine Learning Research, 2026

Long-horizon tool-using tasks sometimes benefit from revisiting earlier subtasks, but revisiting costs routing, coordination and tokens. The paper studies complete cyclic subtask graphs, in which every subtask can hand off to every other and a single router picks each transition by checking natural-language criteria, and compares them with ReAct and dependency-ordered graphs on TextCraft, ALFWorld and Finance-Agent.

Cyclic graphs raise ALFWorld success from 33.8% (ReAct) to 58.2%, and the benchmark-generic graph keeps 87-96% of its TextCraft success when half of its routing decisions are replaced at random. ReAct stays stronger and cheaper on TextCraft; the paper maps where revisiting pays for itself and what it costs.

OpenReview PDF Code

BibTeX
@article{gharzeddine2026cyclic,
  title   = {Complete Cyclic Subtask Graphs for Tool-Using {LLM} Agents: Flexibility, Cost, and Bottlenecks in Long-Horizon Workflows},
  author  = {Gharzeddine, Luay and Saab Jr, Samer},
  journal = {Transactions on Machine Learning Research},
  year    = {2026},
  url     = {https://openreview.net/forum?id=FAkarhXCfI}
}

Submitted

BeatPredict

Luay Gharzeddine (first author)

A leakage-safe ECG dataset and machine-learning and deep-learning benchmarks for predicting arrhythmia in the next heartbeat.

Selected projects

cyclic_task_graphs

The paper's reference implementation and full experiment harness.

A planner builds subtasks and transition criteria; an executor runs the current subtask with tools; a router updates memory and picks the next subtask, looping until the goal is reached or the budget runs out. Planner subtasks + criteria Executor tool calls, local budget Router memory + next subtask segment trace next subtask goal reached or global budget used

A workflow controller in which every subtask can transition to every other, compared against ReAct and dependency-ordered DAGs. Every table and ablation in the paper (about 33,600 episodes across eight experiment families) reproduces from YAML configs, with dry-run cost estimates, spend caps, resumable runs and an analysis CLI that regenerates the paper's tables.

  • 58.2%ALFWorld success, vs 33.8% for ReAct
  • 87-96%of TextCraft success kept under random routing faults
  • 3benchmarks: TextCraft, ALFWorld, Finance-Agent

Repository Paper

text2sql-cfg

Text-to-SQL where the model cannot write an unsafe query.

Terminal recording: a BIRD question answered by the pipeline, then a prompt injection that the model refuses and the grammar rejects

The model can only write SQL through an OpenAI custom tool constrained by a Lark grammar built from the database schema, so writes, stacked statements, comments and unknown columns cannot be sampled at all. Three more layers assume the grammar failed anyway: an AST validator, a read-only sandboxed executor, and tenant scoping injected by the server, so the model never sees a tenant id. Every number reproduces offline from a committed response cache, with no API key.

  • 85.2%held-out accuracy on corrected BIRD Mini-Dev labels (68.0% on original labels)
  • $0.0014per question
  • 0 of 275adversarial attacks reached the database or leaked tenant data

Repository Results Threat model

ditto

Extract a whole website into JSON with about one LLM call per page template.

Chart of LLM cost per 1,000 pages against F1: ditto matches per-page LLM extraction at a small fraction of the cost
LLM cost per 1,000 pages (log scale) against F1, from the repository's benchmark report.

ditto clusters a site's pages by template, shows an LLM one compressed page per template, and gets back reusable XPath specs. It verifies every selector on every page, repairs what fails, and then extracts the whole site with no LLM at all. Benchmarked on SWDE, a Common Crawl silver set and live sites against Crawl4AI, AutoScraper, per-page LLM extraction and hand-written XPath.

  • 0.846F1 on the SWDE test split, vs 0.845 for per-page LLM extraction
  • $0.08per 1,000 pages, vs $22.59
  • ~290xcheaper at the same accuracy

Repository Benchmarks

Experience

  1. Senior AI Engineer

    2026 - present

    X-Tends fintech

    LLM orchestration for a bank's customer-facing assistant: intent routing and specialist agents across 20+ banking workflows, safety controls for regulated actions (RAG grounding, approved tools, audit logs, human review), a multilingual voice assistant, and a read-only, tenant-scoped SQL agent for transaction analytics.

  2. AI Engineer

    2024 - 2026

    Waicah medical AI startup

    Clinical documentation that turns consultations into physician-ready reports (drafting time from 20 minutes to under 5), a fine-tuned report model with evaluation workflows (physician ratings from 3.8 to 4.5 out of 5), a physician-approved intake chatbot, and extraction agents that let physicians review cases 2-3x faster.

  3. AI Engineer, then director-level

    2024

    Tenn.ai

    Multi-agent platform for automating business workflows: an intent-to-workflow engine powering 40+ workflows at a 90%+ completion rate, and an 8x cut in LLM costs through prompt caching, context summarization and complexity-based model routing. Promoted to a director-level role after four months.

  4. BSc Computer Science

    2021

    American University of Beirut

Contact

The best way to reach me is email: luaygharzeddine@gmail.com. I'm also on LinkedIn and GitHub.