Research

Carson Rodrigues

Independent researcher working at the intersection of production AI systems and their measurement — voice-AI latency, LLMOps, the Model Context Protocol, multi-agent reliability, human–AI trust, and clinical ML. 17 papers — 1 accepted at ICANN 2026 (Springer LNCS), with 10 more under peer review and 11 live preprints. Most studies are built on systems I ship in production, so the methods are evaluated under real workloads rather than in isolation.

iD0009-0001-7195-6742 carson@celabe.comAffiliation: Celabe, Research Division

ORCID Google Scholar DBLP arXiv OpenReview OSF / PsyArXiv Research site

Papers

Accepted

Under review

Live preprints

Voice AI systems & latencyLLMOpsModel Context Protocol (MCP)Multi-agent reliabilityHuman–AI trustAgentic safetyClinical ML

Papers & preprints

Updated September 2026

AcceptedICANN 2026·Conference · Springer LNCS

Latency Optimization for a Production Voice AI Platform

Carson RodriguesiD (Celabe), Oysturn VasiD (University of Waterloo)

A systems-level latency study of a production voice-AI platform (Anthropic Claude intent detection + ElevenLabs TTS over a NestJS WebSocket pipeline). The central finding is that running intent detection and TTS concurrently — rather than shaving any single stage — is the highest-leverage optimization, cutting median end-to-end latency from 3,277 ms to 1,909 ms.

−41.8% p50 latency (3,277 → 1,909 ms)Voice AILatencyLLM systems

Accepted at ICANN 2026 (peer-reviewed; Springer LNCS proceedings). Registration confirmed; presented as a printed poster (poster session Thursday 17 September 2026).

SubmittedIEEE Software·Journal · magazine article

MCP Server Architecture Patterns

Carson RodriguesiD (Celabe), Oysturn VasiD (University of Waterloo)

A pattern catalogue for production Model Context Protocol (MCP) servers — a Gamma-format taxonomy, anti-patterns, and cross-cutting concerns — validated with a real inter-rater reliability study (kappa = 0.76). Proposes a scoped Proxy-Aggregator pattern once tool counts exceed a practical threshold.

Inter-rater kappa = 0.76MCPSoftware architectureAgents

arXiv preprint doi:10.48550/arXiv.2606.30317

Condensed magazine version submitted to IEEE Software on 29 June 2026; the extended 9-page version is openly available on arXiv (cs.AI primary, cs.SE cross-list) under the fuller title “MCP Server Architecture Patterns for LLM-Integrated Applications”.

SubmittedAAAI 2027·Conference · double-blind

When Do LLMs Replace Fine-Tuned NLU? A Decision Framework for Intent Detection

Carson RodriguesiD (Celabe), Oysturn VasiD (University of Waterloo)

A decision framework for choosing between LLM classifiers and fine-tuned NLU on noisy production transcripts. Shows that full-data TF-IDF still reaches 95.2% on ATIS, and maps the regimes where an LLM-based intent classifier is — and is not — worth its cost and latency.

95.2% on ATIS (full-data TF-IDF baseline)NLUIntent detectionLLM evaluation

arXiv preprint doi:10.48550/arXiv.2608.20371

Submitted to AAAI-27 on 1 July 2026 (double-blind review; Montréal, February 2027). Preprint openly available on arXiv (cs.CL primary, cs.AI cross-list).

Under review · rebuttal submittedACML 2026·Conference · double-blind

When Is an LLM Worth It for Hyperparameter Optimization? A Budget-Matched Study

Carson RodriguesiD (Celabe), Oysturn VasiD (University of Waterloo), Isaiah Abner DCosta (University of Queensland), Nithish Kumar Prabhakaran (University of Queensland)

A budget-matched, multi-seed study of whether an LLM advisor actually helps hyperparameter optimization on tabular data. The deflationary finding: the advisor's strong first guess is not an LLM output but a fixed default configuration evaluated before any model call. Once classical search (random, Optuna-TPE, Bayesian optimization, successive halving) is seeded with that same default, the apparent lead collapses within a handful of evaluations and the LLM adds no measurable generalization benefit.

LLM warm-start = a default config, not the modelAutoMLHyperparameter optimizationLLMs

arXiv preprint doi:10.48550/arXiv.2606.21641

Submitted to ACML 2026 on 26 June 2026 (Conference Track, double-blind). Four reviews arrived on 31 August 2026; the author rebuttal, submitted 1 September, ran every experiment the reviewers asked for: an exact seeded control for TPE and GP-BO, an 11-point initialisation sweep, a meta-learned initialisation, a 31-task benchmark screen, deep-learning HPO, and a prompt ablation reported in full, including the one result that cuts against the paper's original recommendation. Notification 15 September 2026. Preprint on arXiv (cs.LG · cs.AI); the harness and a script that reproduces every statistic are released.

Resubmitted to IJHCI · preprint liveIJHCI·Journal · review article

Personality Traits and Trust in Large Language Models: A Scoping Review

Carson RodriguesiD (Celabe), Simran Marian Rebello (Celabe)

A scoping review of how personality traits shape user trust in large language models, extended toward agentic (action-taking) systems with a set of testable propositions for future empirical work.

Scoping review + agentic propositionsHuman–AI trustPersonalityScoping review

PsyArXiv preprint doi:10.31234/osf.io/2mdu8

Preprint openly available on PsyArXiv. Resubmitted to the International Journal of Human-Computer Interaction (IJHCI) on 16 July 2026 via a Taylor & Francis transfer, awaiting editor assignment. Novelty: Big Five traits as a moderator of LLM trust calibration.

Preprint · submitted to JAIRJAIR·Journal

Hallucination as Context Drift: Synchronization Protocols for Multi-Agent LLM Systems

Carson RodriguesiD (Celabe)

Reframes a class of multi-agent LLM hallucination as context drift: divergence between agents' internal world-states. Introduces the Context Divergence Score (CDS) and the Shared State Verification Protocol (SSVP). Across two domains on Claude Haiku, naive full-broadcast synchronization backfires by propagating one agent's erroneous state (hallucination rate 0.658 vs. 0.492 for no-sync), while SSVP avoids that failure mode and beats full-broadcast using 58% fewer API calls.

Full-broadcast backfires; SSVP at −58% API callsMulti-agentReliabilityEvaluation

arXiv preprint doi:10.48550/arXiv.2606.21666

Preprint openly available on arXiv. Substantially revised in July 2026 — the contamination effect now replicates across three model families (Claude Haiku, Llama-3.3-70B, Qwen2.5) with multi-judge robustness checks — and submitted to JAIR (Journal of Artificial Intelligence Research); the submission is in administrative processing, not yet under peer review.

Submission package ready · medRxiv preprintJournal of Medical Systems (submission ready)·Journal

Calibrated Next-Procedure Recommendation with Split-Conformal Abstention: A MEPS-Anchored Evaluation of Classical and Large Language Model Designs for General Dentistry

Carson RodriguesiD (Celabe), Steffie Dione Rebello (KLE (co-PI))

A machine-learning approach to predicting dental procedures from patient records, evaluated on a synthetic pipeline benchmark plus a real public-data benchmark (AHRQ MEPS 2023 dental visits), with calibrated abstention for out-of-distribution charts.

MEPS: ECE 0.031 · split-conformal selective risk 51.5% → 42.0% at 48.7% coverage · classical 100 vs LLM 72 correct of 200 (paired p = 0.0003)Clinical MLHealthcare AIPrediction

medRxiv preprint doi:10.64898/2026.05.07.26352635

Carson and Steffie share first authorship (equal contribution). Preprint live on medRxiv (v4, 21 July 2026, with TRIPOD checklists as supplementary material). JMIRx Med declined the submission at the editorial screening stage in July 2026 without sending it to peer review; the paper is being reframed around the reproducible MEPS benchmark for resubmission to a health-informatics journal.

Submitted to IEEE AIEI 2027IEEE AIEI 2027·Conference · IEEE Xplore

Fidelity Assessment for Digital Sketching: Why SSIM Fails on Line Art, and What to Use Instead

Carson RodriguesiD (Celabe), Siddhant Desai (Simplit Software Solutions), Amey Kerkar (Don Bosco College of Engineering), Yeshudas Muttu (Don Bosco College of Engineering), Nihal Kamat (Trellissoft), Ritika Ghanti (Cramer America), Saish Pawar (Oneshield)

LearnSketch is a mobile sketching framework that scores a drawing against an automatically extracted reference outline and returns a numeric score plus a visual error map, with the Structural Similarity Index (SSIM) as that score. Measuring what SSIM returns for drawings known to be bad: over 65 references an empty canvas averages 78.9% and reaches 98.0%, outranking a correct drawing displaced by two pixels on 93.8% of references and a drawing of the wrong reference on 96.9%. The cause is sparsity, and blank-canvas SSIM tracks reference ink density at r = −0.952. Intersection-over-union on dilated edge maps separates acceptable from unacceptable drawings far better (AUC 0.984 against 0.766 for SSIM), is zero for an empty canvas, and costs 0.10 ms against SSIM's 2.76 ms. Learned perceptual metrics (LPIPS 0.995, DISTS 0.994) score marginally higher but still rank a blank canvas above a wrong drawing on 24.6% and 13.8% of references.

Empty canvas scores 78.9% SSIM on average · IoU AUC 0.984 vs SSIM 0.766, at 0.10 msImage similaritySSIMMetric auditSketching

Seven-author collaboration (Carson first author). Submitted on 2 September 2026 to the 2nd IEEE International Conference on AI Engineering and Innovation (AIEI 2027; Bangalore, 21 to 23 January 2027, hybrid; financially co-sponsored by IEEE TEMS, proceedings in IEEE Xplore); notification due 30 October 2026. Earlier versions were not accepted at IC-SIT 2026 and were screened out at ICONAT 2026. The audit harness and the scripts that regenerate every number are public: github.com/rodriguescarson/learnsketch-fidelity-audit (MIT), archived at Zenodo doi:10.5281/zenodo.22237791.

Stage-1 Registered ReportPCI Registered Reports → Peer Community Journal·Registered Report

Personality and Over-Delegation to Agentic LLMs

Carson RodriguesiD (Celabe)

An empirical follow-up to the trust scoping review (Paper 05), testing its propositions on over-delegation to agentic, action-taking LLMs. Stage-1 manuscript complete, with the jsPsych experiment harness built and the design powered at 0.80 for N = 320.

Stage-1 complete · powered at 0.80 for N = 320Agentic safetyHuman–AI trustOver-delegation

Stage-1 Registered Report on the PCI-RR route (Peer Community Journal). PCI-RR is closed to new submissions between 1 July and 1 September 2026, so submission is scheduled for September.

Under review at EAAIEngineering Applications of Artificial Intelligence·Journal · Elsevier

Done Is Not Correct: Measuring Silent Failures and Self-Verification Calibration When LLM Agents Take CAD Actions

Carson RodriguesiD (Celabe), Clive Rodrigues, Aravind Reddy G

A reproducible benchmark for agentic CAD. When an LLM agent writes parametric CAD code from a natural-language spec, syntactic success (the code runs, the geometry renders) decouples from semantic correctness (right dimensions, valid constraints, manufacturable features). The harness measures that silent-failure gap directly and tests whether claim-conditioned self-verification closes it.

468 runs · single-shot 14.1% silent failures (ECE 0.32) → 3.2% with enforced self-checks · expert–grader agreement κ = 0.64 (0.87 excluding one ambiguous task)Agentic safetyReliabilityEvaluation

Benchmark complete (4 models × 3 conditions); independent CAD-engineer validation done (blind pass/fail scoring of 81 parts). Submitted to the ASME Journal of Computing and Information Science in Engineering on 4 August 2026 (JCISE-26-1424); the editor returned it at pre-screening on 10 August without review, as a benchmark rather than a new modelling methodology, after JAIR had declined it on scope the same week as better suited to an application venue. Submitted to Engineering Applications of Artificial Intelligence (Elsevier) on 4 September 2026 after a rigor pass that added a tolerance-band sensitivity analysis and a disclosure of the pre-registered predictions; the benchmark, all attempt records, and the analysis code are public (GitHub, Zenodo doi:10.5281/zenodo.22286324). The arXiv preprint is pending moderation.

Reframed · MARS audit underwayJournal TBD (health informatics)·Journal

Oravira: Design, Privacy-by-Design Architecture, and App-Quality Audit of a Mobile Oral-Health App for People Living with HIV

Carson RodriguesiD (Celabe), Steffie Dione Rebello (KLE)

Oravira is a privacy-by-design mobile app that helps people living with HIV self-manage their oral health. The paper documents the system architecture, the privacy model that keeps sensitive status data off any server, and a MARS-based quality audit of the app-store landscape it sits in — a design-and-architecture contribution that stands on its own, without patient data.

App live on iOS + Android · MARS app-store auditClinical MLHealthcare AIPrivacy

Reframed in July 2026: the clinical evaluation route through the KLE ethics committee and CTRI trial registration is no longer part of this paper, so it needs no ethics approval; a KLE-run study is referenced as future work only. The MARS app-store audit is underway — a two-strand systematic search across both stores (985 candidate apps screened) found no other consumer app at the HIV × oral-health intersection, with 11 apps passing screening for full scoring.

Submitted · decision pendingREALM @ EMNLP 2026 workshop·Workshop · non-archival

DevTwin: Automatic Coverage Overstates How Well a Passively-Formed Memory Knows a Developer

Carson RodriguesiD (Celabe)

Personal-memory systems for language agents are usually scored on retrieval: whether a relevant chunk can be found in a stored history. DevTwin argues this misses what a personal memory is for, which is to know the person rather than the transcript, and benchmarks whether a memory built passively from a developer's own public artifacts (git history, papers, project documents) can answer identity questions about them.

Human raters put identity knowledge at 0.45 where the automatic proxy saturates (paired gap 0.55, n = 10)Agent memoryEvaluationPersonalization

Submitted to REALM @ EMNLP 2026 on 2 July 2026 (non-archival). The decision is still pending as of early September 2026. The human ratings come from ten consented raters on a single developer's memory; the multi-subject study is reserved for the archival version.

Under review at NBDTNeurons, Behavior, Data analysis, and Theory (NBDT)·Journal

Predicted Cortex Is Not a Domain-General Prior: A Matched-Control Audit of Brain-Encoding Features for Video Memorability

Carson RodriguesiD (Celabe)

Do a brain-encoding foundation model's predicted fMRI responses forecast video memorability better than the visual backbone they are built on? Using the TRIBE encoder over its own V-JEPA2 backbone as a matched control, the answer flips with the dataset rather than yielding a single winner. The reversal is tested directly as a dataset-by-representation interaction, not inferred from two separate comparisons. The predicted-brain projection carries a small but real, vision-orthogonal signal that helps on one dataset and not the other, which cautions against treating a brain-encoding readout as a free upgrade over the backbone.

Dataset-dependent reversal, tested as an interaction: +0.097 (95% CI [+0.032, +0.160], p = 0.001) · Memento10k: backbone 0.594 > brain 0.544 · VideoMem: brain 0.415 > backbone 0.368 · complete separation over 10 CV seeds (0/10 and 10/10); cross-dataset transfer inherits the flipNeuro-AIRepresentation learningEvaluation

arXiv preprint doi:10.48550/arXiv.2607.16292

Preprint openly available on arXiv (2607.16292, v4). The derived predicted-response arrays are released for reproducibility with InterDigital's written approval (Zenodo DOI 10.5281/zenodo.21532633 plus a GitHub release); the source datasets and their memorability scores are not redistributed. Results were shared with the dataset authors first, and the VideoMem author replied positively and raised a possible benchmark collaboration for September 2026.

Preprint · Zenodo DOI liveZenodo preprint · Minds and Machines (submission ready)·Perspective · position paper

Hallucination Is Generative Memory With Its Verifier Turned Down: One Constraint Axis Links Dreaming Sleep and LLM Confabulation

Carson RodriguesiD (Celabe)

A perspective paper arguing that dreaming and LLM hallucination are two expressions of one computation: generative reconstruction from distributed memory under incomplete constraints. Neither the sleeping brain nor an autoregressive model retrieves stored records; both synthesise output by recombining learned representations. A four-stage account (encoding, latent representation, generative reconstruction, reality verification) locates the difference between a reliable output and a hallucination or a dream on one ordinal axis: the strength of the constraints and verification acting on the generator, attenuated in dreaming sleep and absent in free decoding. The paper states where the mapping holds and where it breaks, recasts mitigation as restoring verification rather than removing generation, and sets out five falsifiable predictions with their current status. A small pre-registered test on 300 PopQA questions finds the hallucination rate of two language models falling at every step from forced closed-book answering to open-book answering with a verifier (50.7 to 7.0 percent and 46.0 to 5.3 percent), and falling with entity popularity when unconstrained.

Four-stage framework · one ordinal constraint axis · 5 predictions (2 partially tested) · PopQA vignette passed on 2 modelsMemoryHallucinationNeuro-AI

Zenodo preprint doi:10.5281/zenodo.22314160

Sole-authored perspective paper; 83 references independently verified against Crossref and arXiv. Retitled after a September 2026 audit; the constraint axis is now operationalised and tested by a small pre-registered vignette (code and cached responses released). Section 5.4 anchors the controlled-generation argument to Paper 17's measurement.

Under review at DMLR · Zenodo preprintDMLR (Journal of Data-centric Machine Learning Research)·Measurement · platform audit

Silent Failure in Automated Model Adaptation

Carson RodriguesiD (Celabe)

Automated adaptation platforms fine-tune a base model on a user's data and report success through numbers a user cannot verify. Instrumenting one platform across 176 fine-tuning jobs, 86 launches, and 13 datasets through its internal API: the core capability works (augmentation improved the held-out score on 11 of 11 paired seeds, mean +14.4, 95% bootstrap CI 12.4 to 16.4, sign test and Wilcoxon p=0.0010), yet the reported numbers do not separate a good adaptation from a bad one: the quality grade collapses toward a fixed attractor, and the displayed and held-out win rates disagree in sign in three of eight paired comparisons. It defines silent adaptation failure, proposes six machine-checkable audits, and reports a verifier that caught nine failures, four in its own scaffolding.

176 fine-tunes · +14.4 on 11/11 seeds (CI 12.4–16.4, p=0.0010) · 6 auditsData-centric MLMeasurementLLMOpsReliability

Zenodo preprint doi:10.5281/zenodo.21939799

Sole-authored; spun out of a public model-adaptation challenge. The platform is deliberately not named in the paper, and its vendor received the full findings, with job identifiers and reproduction detail, before publication. Preprint published on Zenodo (CC BY 4.0) and submitted to DMLR on 15 August 2026. All external citations verified against arXiv and CrossRef; internal rigor review 2026-08-09.

Under review at DMLR · dataset releasedDMLR (Journal of Data-centric Machine Learning Research)·Benchmark · dataset

Quantifying Chart Deception: Continuous Deception-Severity Estimation for Chart-Reading Models

Carson RodriguesiD (Celabe)

Charts can mislead without containing a single false number: a truncated axis, an inverted scale, or a stretched aspect ratio changes what a reader concludes while every label stays accurate. This paper makes deception a continuous, exactly computable quantity: a counterfactually-augmented dataset generated directly from distortion parameters (2,430 ground-truth rows carrying a 33-field machine-checkable schema, 636 with an exact Tufte lie factor, plus a 2,129-row failure-weighted training projection), a base-model difficulty map showing accuracy collapses exactly where the deception lives (continuous severity, extraction under a distorted axis, mechanism naming), and a held-out Deception Atlas of 20 real published misleading charts verified against primary sources. Training against the exact ground truth with reinforcement learning from verifiable rewards is pre-committed in an analysis plan fixed before any run.

2,430 exact-ground-truth rows · 636 exact lie factors · 20-chart Deception Atlas · RLVR plan fixed in advanceChart understandingVision-language modelsRLVRBenchmarksShortcut learning

Zenodo preprint doi:10.5281/zenodo.21939874

Sole-authored. The dataset, 954 chart renders, and the Deception Atlas are released under CC BY 4.0 (Zenodo DOI 10.5281/zenodo.21939803); the severity target is computed from each chart's construction, so the reward needs no human rater and no judge model. The RLVR training runs follow the fixed plan and will appear in a follow-up paper. Preprint published and submitted to DMLR on 15 August 2026. Companion to Paper 15.

Preprint · Zenodo DOI liveZenodo preprint · arXiv pending·Preprint · interpretability

The Look of a Lie: Chart-Deception Probes on Vision-Language Model Activations Read Distortion Geometry and Are Mechanism-Bound

Carson RodriguesiD (Celabe)

Deception probes on language models answer a binary question: is this response deceptive? Charts allow a graded one, because a chart rendered from its own distortion parameters has an exact Tufte Lie Factor. 1,685 charts are rendered from real statistical series under parameterised distortions and linear probes are trained on the activations of vision-language models viewing them (Qwen2.5-VL 3B, 7B, 72B; Gemma-3-4B). Whether a chart is misleading is decodable at nearly every depth (test AUROC 0.994), but a preregistered hard negative shows what the bit is made of: the same probes flag 68 to 87% of honest charts that merely share a lie's axis geometry. Ridge probes on the exact Lie Factor reach held-out R² of 0.57, 0.61 and 0.74 in deep layers, yet an oracle that knows only which mechanism was applied scores 0.90 on the same split, so graded recovery within a mechanism is untested and the diagnostic that settles it is pre-specified. The probe's Lie Factor error (MAE 0.70) is smaller than the model's own stated estimate (1.06), which is worse than a constant guess (0.76). The direction is mechanism-bound: it transfers across axis-geometry distortions, reverses rank order on a sealed held-out mechanism, and steering with it leaves the model's descriptions unchanged.

Misleading-or-not AUROC 0.994 · hard negative 68–87% · Lie Factor R² 0.74 (72B) below a mechanism oracle at 0.90 · probe MAE 0.70 vs stated 1.06InterpretabilityVision-language modelsChart understandingProbing

Zenodo preprint doi:10.5281/zenodo.22134653

Sole-authored follow-up to Paper 16, built on its PolyChart generator and exact lie factors. A September 2026 audit withdrew the graded-severity claim (a mechanism-identity oracle beats the probe on the test split) and retitled the paper around the hard negative and the mechanism bound; the within-mechanism diagnostic is pre-specified and awaits a re-extraction. Every analysis was fixed before extraction or is dated as an extension.

Several papers are under double-blind review, so author lists and venues may be anonymized in the submitted copies. DOIs are linked where a preprint is publicly posted; full manuscripts, data, and code for any work in progress are available on request — carson@celabe.com. ORCID: 0009-0001-7195-6742.

Available for senior AI / contract / FDE work

Building something with AI?

Voice agents, MCP servers, LLM pipelines, agentic workflows — pick a slot, drop a message, or send your email and I'll reply within a day.

Book a 30-min call WhatsApp me

or leave your email

Replies within ~24 hours · Remote-first · global · open to relocation