Celabe Research Division
The research arm of Celabe. We sit where building and studying AI meet — shipping production systems while researching how those systems behave and how people use them. Our work spans the stack, from the latency budget of a voice pipeline to the psychology of when someone over-trusts a language model, with one aim: research that makes AI faster, more reliable, and safer to rely on.
Celabe Research Division is a small, focused group doing applied AI research in the open. We are practitioners first: we build and operate real systems, and the questions we study come from problems we hit shipping them.
That gives our work a particular character — empirical, systems-aware, and human-centered. We care less about benchmarks in the abstract and more about whether a system is fast enough to use, reliable enough to depend on, and designed so the people using it stay in calibrated control.
Latency, intent detection, and the systems engineering that make voice and chat AI fast and reliable in production.
Design patterns for LLM-integrated software — agent/tool protocols such as MCP, retrieval, and robust structured integration.
Using LLMs to build and optimise ML systems: model selection, hyperparameter tuning, and developer-in-the-loop workflows.
Hallucination, context drift, and synchronisation in multi-agent LLM systems — keeping distributed reasoning coherent.
Trust, reliance, over-reliance, and over-delegation — and the interface and governance design that keeps people in calibrated control.
LLM augmentation built to support professionals, not replace them — including clinical decision support designed for human oversight.
Featured framework
Our central contribution to human–AI trust. The PTC framework models stable individual differences — Big Five traits, dispositional propensity-to-trust, anxiety — as distal moderators that act through proximal mechanisms (anthropomorphic attribution, cognitive offloading) and are bounded by system and context features, to shape when people calibrate trust in AI well — and when they don't.
We extend it from advisory language models to agentic, action-taking systems, where the failure mode shifts from over-reliance on advice to over-delegation of action — and we derive testable propositions and design implications for keeping a human meaningfully in the loop.
We study the systems we build and deploy, and test claims against behaviour — not attitudes alone.
Every citation is checked against canonical scholarly records (CrossRef, arXiv) before release.
We post preprints, share study-level datasets, and use persistent identifiers for everything we publish.
We design for calibrated human control and oversight, not automation for its own sake.
Author manuscripts shared as preprints. We hold ourselves to verifiable scholarship — every reference in our work is checked against canonical scholarly records before release.
C. Rodrigues, S. M. Rebello
A scoping review of 60 studies introducing the Personality–Trust Calibration (PTC) framework, extended to agentic LLMs.
C. Rodrigues et al.
A clinical decision-support recommender that sequences the next dental procedure from the whole patient chart — designed to augment the dentist, surface rationale, and abstain when unsure.
Read preprintWe publish in the open and make our work checkable: persistent identifiers, shared datasets, and a verified reference base for every paper.
We welcome collaboration across applied AI — real-time systems, LLM architecture, multi-agent reliability, and human–AI trust and safety. For research enquiries, partnerships, or data sharing, get in touch.