Skip to content

Research

Work we can point to.

Celaron grew out of research internships at QCRI and QEERI; this page collects the public output.

PEER-REVIEWED · BENCHMARKS · OPEN PLATFORMS

  1. 01

    2026

    Continual learning · Model merging · Evaluation

    Acquisition, Retention, and Fisher-Proxy Scale in Continual Model Merging

    Working manuscript · independent audit of Phan et al. (2025), arXiv:2509.23592

    Built an order-aware 10×10 accuracy-matrix harness and ran a 27-run, three-seed, three-order audit of a published continual-merging pipeline. Uniform α = 1/t cut forgetting from 0.054 to 0.016 (3.4×) but cost 5.57 pp of learning accuracy; decomposed the trade-off and showed that Fisher-proxy scale acts as an implicit merge coefficient.

    Audit code & results

  2. 02

    2025

    Catalysis · Generative models · Interpretability

    CatForge — open catalyst design platform

    QCRI · first author on the system paper (manuscript in submission)

    Built the core synthesis, evaluation, and prediction components of the CatForge catalyst-design platform — Copula/GAN augmentation, XGBoost/RF ensembles, SHAP/LIME interpretability, and a constrained Pareto/TOPSIS optimization studio. Validated across four datasets with 71.8–78.2% aggregate synthetic fidelity and 0.3–6.8 pp agreement with matched lab cases.

    Visit CatForge

  3. 03

    2026

    LLM evaluation · 3D reasoning · Benchmarking

    C2CAD-BENCH — benchmarking LLM spatial reasoning

    QEERI internship project · contributor to implementation, experiments, evaluation, and manuscript (not founder)

    During my QEERI internship, I contributed to a code-free benchmark that tests LLM spatial reasoning through deterministic geometric and structural checks: 75 CAD cases across 25 families, scored on coverage, geometry, and semantic validity. The paper snapshot evaluates 13 models.

    View the benchmark

  4. 04

    2025–26

    MCP · 3D geometry · Manufacturability

    ThinQA — agentic biomimicry design pipeline

    QEERI internship project · MCP-based generation, validation, and rendering (not founder)

    During my QEERI internship, I worked on an LLM-driven biomimicry design pipeline with a three-layer MCP/JSON-RPC architecture for STL validation, GPU-accelerated rendering, and automated feedback — achieving a 100% watertight mesh output rate and under-six-minute generation times for evaluated gyroid-catalyst and vascular heat-sink cases.

    Visit ThinQA

  5. 05

    2026

    Precision oncology · Data integration · Clinical evidence

    CAPRI — precision oncology workbench

    QCRI / HBKU Precision Medicine · A1 research poster

    Built the application and AI layer during a Summer 2026 research internship: a FastAPI / React workbench unifying twenty-nine biomedical sources with variant normalization, AMP/ASCO/CAP Tier I–IV evidence ranking, conflict detection, structure-aware protein–drug interpretation, AutoDock Vina docking, cohort analytics, and de-identified tumor-board reports.

    Open CAPRI

How research becomes product

CatForge and CAPRI began as research projects and now run as deployed platforms. Matflow's benchmark numbers are published and reproducible from bundled datasets rather than asserted. Every claim stays tied to its source — a paper, a benchmark, or a live product.

Read it, then break it.

The papers, benchmarks, and platforms are all public. If something does not hold up, tell us — that is how the work gets better.

bassil@celaron.comSee the platforms