RJL Robert Leggieri

Applied AI · ML engineering · Technology leadership

Technology leader building practical AI solutions.

I build AI
for the
messy middle.

I bring over two decades of systems ownership to applied AI work across language, speech, vision, and enterprise data, backed by an M.S. in Artificial Intelligence from The University of Texas at Austin.

View selected work Download résumé
Email LinkedIn ↗ GitHub ↗

Selected work

Research questions
to working prototypes.

01
Text-Prosody AlignmentCase Studies in Machine Learning · UT Austin

The "I'm Fine" Effect

When the words say one thing and the voice says another.

I designed and evaluated the Alignment Discrepancy Score (ADS), an interpretable measure of the distance between text-derived sentiment and audio-derived prosody in a shared valence-arousal space.

Data189 DAIC-WOZ clinical interviews
Unit of analysisUtterance-level alignment
Evaluation5-fold participant-level cross-validation
Observed result0.62 cross-validated AUC; 0.49 audio-only; 0.70 text-only

What the result showed: ADS did not sharply separate PHQ groups. The PHQ-elevated group instead showed a tighter distribution with fewer high-ADS outliers, consistent with reduced emotional range. A threshold selected for case detection achieved 80% sensitivity; a balanced threshold yielded 74% sensitivity and 51% specificity.

Scope: Sentiment-ADS outperformed the audio-only baseline but trailed the text-only model. Its value is interpretability: ADS exposes cross-modal variability that text-only prediction cannot show. It is best understood as an exploratory cue, not a replacement for text features or a diagnostic tool.

02
BQE CORE AI Data ConnectorProfessional project · Gutschick, Little & Weber, P.A.

A governed data pipeline
from business systems to AI-ready planning.

Built a read-only ETL and reporting connector that turns historical BQE CORE operational and financial data into a normalized, auditable warehouse. It gives downstream AI systems traceable project and workload history while keeping validation and human review explicit.

AccessOAuth 2.0 Authorization Code with PKCE; read:core only
Data engineeringPagination, retries, deduplication, incremental sync
AI readinessHistorical patterns structured for workload planning and decision support

Deployment context: The connector feeds an internal reporting and planning workflow that also uses a proprietary workload workbook. Public examples remain synthetic.

Inspect the public repository
03
PraxiComArchived 2025 prototype · UT Austin NURSING-AI Challenge

Affective simulation for complex clinical conversations.

A voice-based simulated patient for emotionally demanding communication practice.

PraxiCom student interface with a synthetic patient profile and faculty chat
Student conversation interface
PraxiCom faculty dashboard showing patient-centered communication criteria with Likert-scale ratings and notes
Faculty evaluation dashboard

PraxiCom, formerly CareSpeak, was an individual prototype developed with guidance from UT Austin nursing faculty. I built the student and faculty interfaces, real-time session relay, structured patient personas, and evaluation workflow; Hume EVI supplied the low-latency, natural voice-to-voice layer used for rapid prototyping.

StackReact · Node.js · Socket.IO · YAML · Hume EVI
Persona & evaluationGuideLLM-informed goals and context · live faculty monitoring · Likert ratings
BoundarySynthetic patients; no PHI; not clinically validated

Engineering & design: Hume EVI resolved the prototype's immediate latency constraint, while GuideLLM-informed prompting kept faculty-defined goals, topics, and conversational milestones in context. Future work would use a more modular speech pipeline to allow finer prompt control between stages.

Read the public case study Watch the demo

Experience & education

AI work grounded in operating reality.

More than two decades of responsibility for systems people rely on.

2004-Present

Network Manager

Gutschick, Little & Weber, P.A.

Own technology operations, security, automation, vendor strategy, and long-range infrastructure planning for an engineering firm supporting more than 40 users. Translate cross-functional requirements into dependable, cost-conscious systems.

M.S., Artificial Intelligence

The University of Texas at Austin · GPA 4.0

B.S., Data Science

University of Maryland Global Campus · Summa Cum Laude

Selected tools
PythonPyTorchscikit-learnHugging FaceOpenCVSQL / SQLitePowerShellETL pipelines

What carries forward Systems judgment · stakeholder fluency · security mindset · operational ownership · pragmatic delivery

Graduate coursework · Advances in Deep Learning · UT Austin

Modern model building,
under real constraints.

Private course repositories; methods and implementation scope are summarized here.

01

Memory-efficient fine-tuning

Implemented mixed precision, LoRA, 4-bit quantization, and QLoRA to compare memory-efficient fine-tuning approaches.

Mixed precisionLoRA4-bitQLoRA
02

Generative image compression

Built patch-level autoencoders, Binary Spherical Quantization, and autoregressive models to explore compact latent image representations.

AutoencodersBSQLatent tokensAutoregressive
03

LLM mathematical reasoning

Implemented structured prompting, custom generation, and parameter-efficient adaptation for mathematical reasoning in smaller language models.

PromptingGenerationPEFTEvaluation
04

Vision-language QA with synthetic data

Built a SuperTuxKart pipeline that programmatically generated nearly one million grounded question-answer pairs for object, count, track, and ego-centric spatial questions.

SuperTuxKart race frame with green and red bounding boxes around karts

Graduate coursework · Deep Learning · UT Austin

Three routes to autonomous driving.

Implemented three SuperTuxKart planners: an MLP mapping ground-truth lane boundaries to future waypoints, a Perceiver-inspired transformer using learned queries and cross-attention, and a CNN predicting vehicle-frame waypoints directly from images.

MLP baseline Learned queries Cross-attention Image-to-waypoint CNN
SuperTuxKart MLP planner navigating the Lighthouse track beside a visualization of lane boundaries and predicted waypoints
MLP lane-boundary planner navigating the Lighthouse track.

Graduate coursework · Natural Language Processing · UT Austin

Training models to learn
the task, not the shortcut.

NLP Final ProjectDeBERTa-v3-base · MNLI · HANS

Dataset cartography for robust NLI

Applied established dataset-cartography methods to map 392,702 MNLI examples across five epochs, then built a custom category-weighted cross-entropy loss that scaled each sample’s contribution according to its easy-to-learn, ambiguous, or hard-to-learn assignment. A custom Hugging Face callback captured per-example training dynamics, while a weighted PyTorch dataset and Trainer override carried those weights into fine-tuning.

The strongest configuration assigned weights of 0.5, 1.0, and 0.5 to easy, ambiguous, and hard examples. It improved HANS adversarial accuracy from 77.49% to 78.59% while retaining 88.47% MNLI validation accuracy. Up-weighting hard examples consistently reduced performance, suggesting label-quality problems in that region.

Download the course report
Data map of 392,702 MNLI training examples plotted by confidence and variability, with easy-to-learn, hard-to-learn, and ambiguous regions labeled
Training dynamics across five epochs reveal distinct easy-to-learn, hard-to-learn, and ambiguous regions.

Additional graduate foundations

The mathematical, decision-making, and governance foundations behind the applied work.

Optimization & online learning

Studied convex programs, duality, optimality conditions, first- and second-order methods, stochastic optimization, online learning, and bandit algorithms across theory- and implementation-focused courses.

Reinforcement learning

Studied value-based, policy-gradient, actor-critic, model-based, and model-free methods for sequential decision-making, including on-policy and off-policy learning and exploration.

Planning & reasoning under uncertainty

Defined symbolic and continuous planning problems and examined heuristic search, adversarial planning, Bayesian state estimation, MDPs, and POMDPs.

Ethics & responsible AI

Applied global ethical perspectives and value-sensitive design principles to fairness, accountability, transparency, power, agency, and consequential AI deployment.

About

AI that earns trust by showing its work.

I’m Robert “B.J.” Leggieri, a technology leader applying systems judgment to natural language processing, speech, computer vision, enterprise data, and human-centered system design.

I’m drawn to problems where predictive performance is only part of the story, and where interpretability, security, educational value, or deployment constraints shape what a useful system should be.

Open to conversations about applied AI and ML engineering, AI solutions, multimodal systems, and responsible AI.

Email me LinkedIn ↗ Download résumé