Real Problems, Real Systems

Active client engagements and a track record of building production AI systems for healthcare and regulated environments.

Active · OngoingMedical Device Intelligence · ECRI Institute

Medical Device Intelligence Platform

ECRI · Global Nonprofit Health Research Organization

Five production pipelines over a 10,422-device catalog: Horizon Scanning (72+ sources, 5,000+ weekly signals → ~100 for analyst review, 0 false negatives), Time-to-Market (~87% accuracy, 77 improvement phases), Attribute Validation (+12.1pp coverage gain), EMBASE Novelty Mining (90.1% accuracy), and FDA Post-Market AI Surveillance — under active FDA BAA contract FDABAA-26-00123.

10,422
Devices in Catalog
98%
Signal Auto-Rejection · 0 False Negatives
~87%
Time-to-Market Accuracy (SME-validated)
Active
FDA BAA Contract FDABAA-26-00123
Read the full case study →
Active · OngoingMedical Evidence AI · Veteran Disability Claims

Veteran Disability AI Systems

Trajector · Senior Lead Data Scientist

5 production AI systems for veteran disability claims: Direct Nexus (99.4% source traceability, 80% LLM judge agreement), Secondary Nexus (22,000+ condition pairs, 70% validity at STRONG confidence), Presumptive Nexus (20 deterministic rules, 0 false positives, confidence 1.0), VA Rating (80–100% exact match, 157 unit tests), and OL Observability (11 automated QA checks, 9-veteran golden cohort, 1.37M+ VA evidence documents). 253+ tests across the platform.

Medical Evidence AI38 CFR Regulatory Logic99.4% Source TraceabilityDeterministic Guardrails253+ Tests

Founder's Prior Enterprise Career & Published Research

Work done by Jorge Rivero before founding PredictIQ AI — at Oracle Corporation and in academic research. Included to demonstrate depth in the domains we serve.

Hospitality AI · Prior Career at Oracle

AI-Powered Room Assignment Optimization

Hospitality / Enterprise

Context

Hotel operators face a daily challenge: assigning incoming reservations to available rooms in a way that balances operational efficiency with guest satisfaction. Traditional rule-based systems produced suboptimal assignments — higher costs, uneven room wear, and unmet guest preferences. Complexity grows exponentially across multi-day planning horizons.

Challenge

Build an automated room assignment system that could model competing objectives — guest satisfaction, operational efficiency, room utilization — and solve the resulting NP-hard optimization problem at enterprise scale, handling thousands of reservations in real time without expensive classical solvers.

Solution

Co-invented a patented AI system that models guest importance (amenities, accessibility, loyalty tier, booking channel) alongside operational constraints (upgrade minimization, room wear balancing, gap reduction). Applied Hungarian method heuristics, greedy assignment, and Lagrangean relaxation to achieve near-optimal solutions at a fraction of the computational cost of classical MILP solvers.

Outcome

Automated assignment of thousands of reservations in under a minute with results comparable to mathematically optimal solutions. Reduced reliance on expensive third-party optimization libraries and provided configurable trade-offs between guest satisfaction and hotel efficiency. Deployed as part of Oracle OPERA Cloud — co-inventor on the related US patent, assigned to Oracle Corporation.

Why This Matters for PredictIQ AI

The same principles — multi-objective optimization, scalable AI under real-world constraints, human-configurable trade-offs — apply directly to how PredictIQ AI approaches complex resource allocation in healthcare (device evaluation pipelines, analyst capacity) and enterprise AI (prioritization frameworks, data routing). Hard problems in one domain rarely stay there.

Combinatorial OptimizationHeuristic SearchLagrangean RelaxationMulti-Objective AIProduction ML
Hospitality AI · Prior Career at Oracle

AI-Based Hotel Demand Modeling

Hospitality / Enterprise

Context

Hotels face competitive pressure to move beyond one-size-fits-all pricing. Traditional demand forecasting assumes all customers behave the same — ignoring differences in willingness-to-pay, booking channel, group size, stay length, and travel purpose. The result: missed revenue and impersonal guest experiences.

Challenge

Build demand models that capture heterogeneous customer behavior at scale — grouping guests into evolving personas and predicting how each persona responds to different room types, rate codes, and prices, while accounting for the influence of offer display order on booking decisions.

Solution

Co-invented a patented system combining dynamic ML clustering (random forest, soft clustering, Gaussian mixtures) with discrete choice modeling (Multinomial Logit) per cluster. An iterative Expectation-Maximization algorithm refines clusters over time as new booking data arrives. The model captures both price sensitivity and display-order effects — a critical factor in how guests actually choose.

Outcome

Achieved 4% improvement in prediction accuracy over static single-cluster models. Enabled personalized recommendations, pricing, and display ordering aligned to guest segments. Provided a foundation for revenue optimization across hotel portfolios. Co-inventor on the related US patent, assigned to Oracle Corporation.

Why This Matters for PredictIQ AI

Combining ML clustering with choice modeling to understand heterogeneous behavior is a pattern PredictIQ AI applies directly — in healthcare, different analyst workflows and device categories require segmented models, not one-size-fits-all pipelines. The same EM-based refinement loop that improved hotel demand predictions drives how our systems improve coverage over successive pipeline runs.

Dynamic ClusteringDiscrete Choice ModelingExpectation-MaximizationRandom ForestRevenue Optimization
Hospitality AI · Prior Career at Oracle

AI-Driven Upsell Modeling with Federated Learning

Hospitality / Enterprise

Context

Upselling is one of the highest-leverage revenue levers in hospitality. Traditional strategies were static — every guest saw the same promotions regardless of preferences or context — leading to poor adoption rates, frustrated customers, and missed revenue.

Challenge

Build personalized, real-time upsell models that account for guest personas, price sensitivity, and display-order effects, while solving two hard constraints: hotels can't share raw customer data across properties, and models must remain interpretable enough for revenue managers to trust and act on.

Solution

Co-invented a patented system combining NLP feature extraction from room descriptions, regularized logistic regression (Elastic Net) for feature selection, and Multinomial Logit choice modeling per guest cluster. Extended the framework with hierarchical Bayesian models to share statistical strength across properties and federated learning to train global models without moving raw customer data — one of the first applications of federated learning in hospitality AI.

Outcome

Personalized real-time upsell offers that adapted to each guest's context. Improved prediction accuracy and stability versus static rules. Quantified willingness-to-pay insights per feature and upgrade type. Co-inventor on related US patents assigned to Oracle Corporation.

Why This Matters for PredictIQ AI

Federated learning — training across distributed data without centralizing it — is a pattern directly applicable to PredictIQ AI's healthcare work, where data governance and privacy constraints are similarly non-negotiable. The hierarchical Bayesian approach to sharing strength across heterogeneous nodes (hotels → hospital systems, device categories) is a core technique in how we build models that generalize without overfitting to any single data source.

Federated LearningHierarchical Bayesian ModelsMultinomial LogitElastic NetNLP Feature Extraction
Hospitality AI · Prior Career at Oracle

Data-Driven Revenue Management for Upgrades

Hospitality / Enterprise

Context

Upgrades and add-ons are a critical revenue lever in hospitality, but traditional pricing models assume customers make a single independent choice. In reality, upgrade selections are interdependent — a guest choosing a room upgrade is more or less likely to choose other add-ons depending on context, loyalty status, and what else is on offer.

Challenge

Build an upgrade optimization framework that models simultaneous customer choices (substitution and complement effects), personalizes offers by segment, and solves the resulting combinatorial pricing and assortment problem in real time at enterprise scale — without sacrificing interpretability for revenue managers.

Solution

Co-developed a framework combining multivariate demand modeling (capturing joint choice behavior), behavioral segmentation by loyalty status and booking context, and real-time assortment and pricing optimization. Transformed the underlying NP-hard combinatorial problem into a near-optimal form solvable in milliseconds — making it deployable in live booking flows.

Outcome

Improved prediction accuracy of upgrade selections versus traditional single-choice models. Delivered quantified willingness-to-pay insights per feature. Enabled dynamically tailored upgrade offerings with measurable revenue gains. Co-investigator on related research conducted in collaboration with academic partners; associated IP assigned to Oracle Corporation.

Why This Matters for PredictIQ AI

Multivariate demand modeling — where choices influence each other — maps directly to how PredictIQ AI structures multi-step AI pipelines: each stage conditions on prior outputs, and optimizing the full chain requires modeling those interdependencies explicitly. The real-time assortment optimization techniques developed here inform how we design decision frameworks that remain tractable under enterprise-scale constraints.

Multivariate Demand ModelingAssortment OptimizationBehavioral SegmentationCombinatorial OptimizationReal-Time Pricing
Hospitality AI · Prior Career at Oracle

ML-Optimized Overbooking & Paid Upgrades

Hospitality / Enterprise

Context

Hotels routinely overbook standard rooms to hedge against cancellations and no-shows, but doing so risks costly walk-offs or excessive free upgrades. Meanwhile, premium rooms sit under-utilized. The challenge: set optimal overbooking limits while monetizing premium inventory through calibrated paid upgrades — across single and multi-day stays.

Challenge

Build a system that learns from historical reservation data to balance revenue lift against denied-service risk, price paid upgrades dynamically, and scale to multi-day admission control — all while producing controls interpretable enough for revenue managers to trust and deploy at check-in.

Solution

Co-invented a framework that estimates upgrade acceptance curves (linear and log-linear decay via maximum likelihood), derives cancellation and occupancy distributions from data, and applies a marginal revenue / marginal loss stopping rule to set optimal overbooking limits. Introduced Dynamic Run-Out Pricing (DROP) — an interpretable, near-optimal policy that adjusts upgrade price in real time using live inventory and expected arrivals. Extended to multi-day stays by evaluating total-stay marginal revenue rather than per-night caps.

Outcome

Data-driven booking limits that reduce denied-service risk while capturing revenue from underused premium inventory. Interpretable upgrade pricing (SCP/DROP) deployable at pre-arrival and check-in. Scalable to enterprise CRS/PMS environments. Co-inventor on the related US patent, assigned to Oracle Corporation.

Why This Matters for PredictIQ AI

The marginal-trade-off framework at the core of this work — learn distributions from data, derive an optimality condition, ship an interpretable control — is the same philosophy PredictIQ AI applies to AI pipeline design. In healthcare, every pipeline decision (when to escalate to a more expensive model, when to trust a cached result) is a marginal revenue vs. marginal cost calculation. Interpretability isn't a nice-to-have; it's what makes the system trustworthy enough to act on.

Revenue OptimizationMaximum Likelihood EstimationDynamic PricingAdmission ControlStochastic Modeling
Hospitality AI · Prior Career at Oracle

Causal ML for Automated Room Pricing

Hospitality / Enterprise

Context

Pricing hotel rooms is a moving target — demand shifts with seasonality, lead time, occupancy, and competitive context. Purely predictive models fail here because pricing asks a causal question: what would demand be if we set price to p? Not just: what was demand when price happened to be p?

Challenge

Build a pricing engine that correctly isolates the causal effect of price on demand (blocking confounders like seasonality and booked-so-far), embeds that estimate in a revenue-maximizing optimizer, and closes the loop so deployed prices generate interventional data that continually refines the model — all at enterprise scale across single and multi-property chains.

Solution

Co-invented a causal pricing engine formalizing a DAG for the demand-generating process and using Double/Debiased Machine Learning (DoubleML) with a log-linear demand form to estimate heterogeneous price sensitivity by room category, booking window, and season. The learned price→demand mapping feeds a revenue optimizer with guardrails for stability. Supports cross-price effects across room types. Designed as a closed-loop system where deployed prices create interventional data that refines both the causal model and optimizer over time.

Outcome

Interpretable pricing recommendations that separate true price sensitivity from seasonality and other confounders. Near-optimal prices under real business constraints, with guardrails preventing large swings while still exploring for learning. Scalable to chain-wide deployment via PMS/CRS API integration. Co-inventor on the related US patent, assigned to Oracle Corporation.

Why This Matters for PredictIQ AI

Causal inference — not just prediction — is a core technique in how PredictIQ AI approaches AI system design. In healthcare pipelines, the question is never just "what did the model predict?" but "what would coverage be if we changed this threshold or data source?" DoubleML and the closed-loop learning mechanics developed here apply directly to how we instrument pipelines for continuous improvement without confounding pipeline changes with model changes.

Causal InferenceDoubleMLHeterogeneous Treatment EffectsRevenue OptimizationClosed-Loop Learning

Ready to build your case study?

Tell us about your problem. We'll help you figure out whether AI is the right solution — and how to measure success from day one.

Book a Strategy Call →