Diego Armando Salinas Lugo

Professional Summary

AI & Machine Learning Engineer, MSc Artificial Intelligence (Distinction), who ships complete systems — data pipeline, model, API, interface, cloud — not notebooks. Builds generative AI at SAP to automate technical fault reporting and founded ARMATUS, a production iOS coach. Delivered predictive maintenance at 16.74-cycle RMSE and R² 0.825 on NASA C-MAPSS via FastAPI and a LangGraph multi-agent layer on Cloud Run. Monterrey, Mexico; TN visa eligible for United States roles.

Technical Skills

Languages

Python, SQL, TypeScript, JavaScript, MATLAB, C

Machine Learning

scikit-learn, XGBoost, TensorFlow, Keras, Random Forest, Gradient Boosting, Support Vector Regression, Neural Networks, Feature Engineering, Hyperparameter Tuning, Cross-Validation (LOPO, LOSO, k-fold)

Generative AI & LLM

SAP Business AI, SAP AI Core, SAP AI Launchpad, SAP Generative AI Hub, OpenAI GPT-4o, LangChain, LangGraph, Retrieval-Augmented Generation (RAG), Vector Databases (ChromaDB), Embeddings, Prompt Engineering, Multi-Agent Systems

MLOps & Deployment

FastAPI, REST APIs, Streamlit, Flask, Docker, Google Cloud Run, MLflow, Git, GitHub, Vercel, React Native (Expo), PostgreSQL, SQLAlchemy, Alembic, Supabase

Data, Evaluation & Domain

Pandas, NumPy, SciPy, NeuroKit2, Signal Processing (Butterworth, Welch PSD, PPG/HRV, vibration), Time-Series Forecasting, SHAP, LIME, ROC-AUC, Precision / Recall / F1, RMSE / MAE / R², Fairness Analysis, Predictive Maintenance, Healthcare & Clinical ML, Affective Computing, Robotics (PID, Fuzzy Logic), Industrial Automation

Spoken Languages

Spanish (native), English (professional), German (basic)

Professional Experience

1. Machine Learning & AI Engineer, SAP

   Jan 2026 — Present

   San Pedro Garza García, NL, Mexico

   • •Design and deploy generative AI services that automate technical fault reporting, improving incident categorization and cutting engineer triage time.
   • •Build LLM-backed extensions on SAP Business AI and BTP, integrating foundation models via SAP AI Core, AI Launchpad and the generative AI hub.
   • •Convert unstructured fault text into structured, routable records, replacing manual classification with a model-driven workflow. Certified SAP Generative AI Developer (2026) on this stack.

2. Founder & Machine Learning Engineer, ARMATUS (armatus.app)

   May 2026 — Present

   Monterrey, Mexico

   • •Founded and shipped a production iOS generative-AI training coach: onboarding, plan generation, live session tracking and progression, bilingual EN/ES.
   • •Architected a closed five-stage protocol (calibrate, plan, adapt, execute, evolve) that regenerates each week from logged sets, PRs and accumulated fatigue rather than serving a static template.
   • •Built a streaming API converting an athlete profile plus daily readiness signals (energy, sleep, time) into a dosed weekly plan with explicit rationale.
   • •Modelled endurance work as real systemic load so strength volume auto-adjusts behind a readiness ceiling; own the full surface from mobile client to API.

3. Data & AI Trainee, Interius

   Sep 2025 — Dec 2025

   San Pedro Garza García, NL, Mexico

   • •Delivered AI MVPs end to end — problem framing, feature engineering, model training, evaluation and deployment.
   • •Built contact segmentation across engagement, geography and channel activity, converting a flat list into targetable cohorts for personalized outreach.
   • •Shipped workload and efficiency diagnostics as containerized Streamlit services on Google Cloud Run for self-serve stakeholder access.

4. AI Researcher, University of Essex

   Nov 2024 — Sep 2025

   Colchester, United Kingdom

   • •Built an end-to-end emotion-recognition pipeline from PPG across 62 participants, reaching Pearson r = 0.5975 and RMSE 0.154 for continuous arousal, above the r ≈ 0.40 reported for comparable PPG-only work.
   • •Engineered 30+ HRV features across time, frequency (Welch PSD) and non-linear Poincaré domains with per-subject baseline normalization.
   • •Validated under leave-one-participant-out cross-validation with nested search, benchmarking Random Forest, XGBoost and SVR to prove generalization to unseen subjects.

Selected Projects

1. Predictive Maintenance Platform — Remaining Useful Life

   rul-dashboard-368785016309.us-central1.run.app

   XGBoost, SHAP, FastAPI, Streamlit, LangGraph, OpenAI, MLflow, Docker, Google Cloud Run

   • •Predicted turbofan remaining useful life on NASA C-MAPSS at 16.74-cycle test RMSE, 12.33 MAE and R² 0.825, with SHAP exposing the driving sensor per prediction.
   • •Served it via FastAPI behind a Streamlit fleet dashboard, plus a LangGraph multi-agent layer answering natural-language questions and emitting PDF briefings.

2. Personal Research Assistant — Retrieval-Augmented Generation

   github.com/DArmandoSalinas/RAG

   FastAPI, LangChain, ChromaDB, OpenAI GPT-4o, text-embedding-3-small, Streamlit, PyPDF

   • •Built document question-answering that chunks, embeds and retrieves via MultiQueryRetriever, grounding every GPT-4o answer in ChromaDB and returning its source snippet.
   • •Indexed a 149-segment corpus and eliminated unsourced answers by making citation a hard requirement of the response contract.

3. Cardiovascular Risk Classification — Leak-Free Clinical ML

   github.com/DArmandoSalinas/HeartDisease-Predictor

   scikit-learn, HistGradientBoosting, SHAP, LIME, ColumnTransformer

   • •Classified coronary disease across 920 patients from four sites at 0.90 test ROC-AUC, 87.3% recall and 0.85 F1, tuning the threshold for recall since a missed case costs more than a false alarm.
   • •Prevented leakage by fitting the ColumnTransformer inside cross-validation, and exposed a transportability gap via Leave-One-Site-Out validation (mean AUC 0.79, worst site 0.70).

4. COAZON — Heart-Health Companion (in development)

   FastAPI, SQLAlchemy 2.0, PostgreSQL, Alembic, JWT, React Native (Expo), TypeScript, OpenAI, Railway

   • •Building a bilingual cardiac-care app that turns a prescription into a living plan — medication reminders, guided blood-pressure logging with clinical-zone trend charts, a deterministic red-flag alarm engine and a doctor-ready PDF report.
   • •Bounded the AI layer so it structures and summarizes but never diagnoses, prescribes or alters a dose, and made it degrade to fully manual without an API key so the app never depends on the model being available.

5. Motor Performance Monitoring — Industrial IoT (1st place, Tec Expo Ingenierías)

   github.com/DArmandoSalinas/Motor-performance-prediction

   Python, Streamlit, signal processing, serial telemetry, statistical thresholding

   • •Won first place at Tec Expo Ingenierías by re-engineering an existing plant maintenance procedure into a live monitoring system, profiling a Sumitomo three-phase motor across five speed baselines and a Haas Mini Mill from combined trajectories.
   • •Scored machine health 0–100 from threaded serial vibration and temperature telemetry with machine-specific adaptive thresholds, keeping the design fully rule-based so every alarm traces to the statistic that fired it; recognized by Rockwell Automation for the integration work.

6. ADHD Screening Support with Fairness Analysis

   github.com/DArmandoSalinas/Predicting-ADHD-sex

   scikit-learn, XGBoost, Keras, KernelPCA, SHAP, LIME

   • •Predicted ADHD outcomes for 1,213 participants from fMRI connectomes and psychosocial data at 82% accuracy, 0.88–0.90 recall and 0.866 AUC; reported results by sex, surfacing a female-recall gap tied to documented underdiagnosis.

Education

1. MSc Artificial Intelligence — Distinction

   2024 — 2025

   University of Essex, United Kingdom

   • •Modules: Machine Learning; Intelligent Systems and Robotics; Data Science and Decision Making; Neural Networks; Natural Language Engineering. Thesis: heart-rate-variability emotion recognition from PPG signals.

2. B.Sc. Mechatronics Engineering

   2021 — 2025

   Tecnológico de Monterrey, Mexico

   • •Coursework: control systems, industrial automation, robot design, manufacturing.