Carson Rodrigues.

Latency Optimization in Production Voice AI Pipelines: A Systems Approach

Published · ICANN 2026

−41.8% p50 latency (3,277 → 1,909 ms)

Labeled Examples Decide the Gap Between Prompted LLMs and Fine-Tuned Intent Encoders: An Audit of Five LLMs and Three Decision Models

Submitted · ACL Rolling Review (target: ACL 2027)

Zero-shot LLMs trail a tuned encoder by 13.4–27.5 pts on ATIS

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

Resubmitted to IJHCI · IJHCI

Scoping review + agentic propositions

A Zero-Shot Decision Model (Jev) Matches a Trained Classifier but Ranks Its Own Errors Worse: Calibrated, Split-Conformal Next-Procedure Recommendation in General Dentistry

Submitted · Journal of Medical Systems

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)

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

Submitted to IEEE AIEI 2027 · IEEE AIEI 2027

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

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

Under review at EAAI · Engineering Applications of Artificial Intelligence

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)

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

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

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 flip

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

Submitted · Submitted to Minds and Machines (Springer)

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

Auditing Hosted LLM Fine-Tuning From the User's Side: Replicates and Answer Keys Show a Service's Default Score Misorders Runs and Its Data Rewriting Breaks Correct Training Targets

Proposal submitted · Proposal submitted to Harvard Data Science Review

481 fine-tuning jobs · default score misorders 44.2% of run pairs · rewriting broke 8.2–19.6% of correct targets

Lie Factors by Construction: Exact Deception-Severity Labels for Parameterised Chart Distortions, and an Inverted Axis That Reverses Three of Ten Models' Readings

Submitted · Submitted to PacificVis 2027

2,430 exact-ground-truth rows · 636 exact lie factors · inverted axis reverses 3 of 10 models (210 pairs, pre-registered)

Chart-Deception Probes in Vision-Language Models Rank Synthetic Axis Exaggeration but Do Not Reliably Beat a Mechanism-Only Baseline, and Qwen Probes Reverse on Real Charts

Submitted · Submitted to IEEE TVCG

AUROC 0.99 on generated charts · no clear gain over a mechanism-only baseline · real Misviz charts: Qwen 7B/72B AUROC 0.43/0.24

MCP Server Architecture Patterns

Preprint · arXiv

Inter-rater kappa = 0.76

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

Preprint · arXiv (cs.CL)

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

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

Preprint · arXiv (cs.AI)

Headline effect withdrawn after a pre-registered re-score

Hyperparameter Advisors Built on Three LLMs Did Not Beat Default-Seeded TPE on Small Tabular Tasks: Four of 15 Pre-Specified Hypothesis Tests Passed, and the Comparison Depends on TPE's Start-Up Trials

Revising · Revising

No LLM advisor beat default-seeded TPE · 4 of 15 tests passed

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

Revising for resubmission · Next venue to be decided; REALM @ EMNLP 2026 workshop declined

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

Personality and Over-Delegation to Agentic LLMs

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

Stage-1 complete · preregistered N = 500

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

In preparation · Journal TBD (health informatics)

App live on iOS + Android · MARS app-store audit

When Frozen-Opponent Replays Overstate: An Audited Study of Opponent-Conditioned Policy Switching in a Two-Player Farming Simulation

In preparation · In preparation

Replayed wins 46.9% → 57.4%; most of the gain gone live

Effect of Implant Depth and Angulation on the Accuracy of Intraoral Scans Made With Scan Bodies and Scannable Healing Abutments: An In Vitro Study

In preparation · In preparation

2 depths × 2 angulations × 2 scan components

ORCID 0009-0001-7195-6742