Skip to content
Sourena KhanzadehAI Research Engineer

Case studies

Systems built to answer a hard question — and to show the answer.

Each case study states the problem, what I personally built, how it was evaluated, what the result was, and where it breaks down. The limitations are on the page, not in a footnote.

  • Preprint + public research artifact2026

    Project Ariadne

    An agent's written reasoning is not evidence that the reasoning caused the answer.

    My role: Designed the intervention protocol, built the auditing harness, and ran the evaluation.

    • Counterfactual replay
    • Trajectory logging
    • LLM evaluation
    • Python

    23 of 30

    audited trajectories showed a faithfulness violation under the counterfactual-intervention protocol

    This is an audit finding about the agents under test, not a failure rate of the tool. 23/30 is 76.7%.

    Read the Project Ariadne case studyPaper for Project AriadneCode for Project Ariadne

  • Prototype architecture and case study2025

    AgentMesh

    A single LLM asked to build software stops at the first thing it gets wrong.

    My role: Sole author. Designed and implemented the framework.

    • Multi-agent systems
    • Tool use
    • Self-repair
    • Python

    Self-repair

    execution feedback drives automatic repair — demonstrated end to end, not benchmarked

    No benchmark, baseline, or success rate is reported in the paper, so no comparative claim is made here.

    Read the AgentMesh case studyPaper for AgentMesh

  • Peer-reviewed research2023–2026

    Search & planning

    When a heuristic gives a search no gradient to follow, the search stalls in a region it cannot see its way out of.

    My role: Co-developed the restarting-random-walk planner (AAAI 2026). Lead author on the separate refactoring-heuristics work.

    • Heuristic search
    • AI planning
    • PDDL
    • Admissible heuristics

    70.3%

    of unbounded-region tasks, where the restarting-random-walk planner was faster than or uniquely successful against breadth-first search

    The 70.3% is measured on unbounded-region tasks specifically, not on planning benchmarks as a whole. "Uniquely successful" means the comparator did not solve the task at all.

    Read the Search & planning case studyPaper for Search & planningPaper for Search & planning

  • Peer-reviewed research2024

    GANsemble

    A 210-image, severely imbalanced dataset is too small to train a classifier that generalizes.

    My role: Co-developed the augmentation search and the class-conditional GAN pipeline.

    • Conditional GAN
    • Synthetic data
    • Deep learning
    • Augmentation search

    91.5%

    downstream classification accuracy — 4 points over duplication-based oversampling, 5.5 over none

    This is downstream classifier accuracy on the microplastics dataset, not a measure of the generated images themselves.

    Read the GANsemble case studyPaper for GANsemblePaper for GANsemble

Building an agent that must work outside the demo?

I work on evaluation, grounded tool use, failure recovery, and auditable agent workflows.

Discuss an agent problem