17 projects shown for All.
ARMATUS
Founder · Production · NLP / LLMs
An iPhone coach that architects the training week — not a template you tap through.
ARMATUS is a consumer iOS product built around a closed protocol: Calibrate → Plan → Adapt → Execute → Evolve. Deep onboarding (goals, equipment, injuries, schedule, lifestyle, external sports) becomes the athlete model. Each week is generated with an overload map and an explicit rationale — then the daily coach doses the session against energy, sleep, time and a readiness ceiling, so the plan yields to capacity rather than ego. The tracker logs sets against a technique standard, with sketches for setup, breathing and common errors. Journey stores PRs, volume and hybrid-sport load as the dataset the next week is regenerated from. Running, cycling, swimming and triathlon enter the algorithm as real fatigue, not a note in the margin. Spanish and English from day one. Production mobile app plus a streaming API for structured plan generation.
iOS shipped app5-step closed protocolES / EN bilingual
TypeScript · iOS · Generative AI · Streaming API · Design system
COAZON
Founder · Healthcare AI · NLP / LLMs
A cardiologist's prescription, turned into a care plan a family can actually follow.
COAZON turns a doctor's prescription into a living household plan: flexible medication reminders that tolerate a real day, a guided blood-pressure and vitals diary with trend charts marked by clinical zone, a deterministic red-flag alarm engine wired to an emergency screen, and a doctor-ready PDF for the next appointment. The AI layer is deliberately bounded — it reads prescriptions, accepts natural-language logging and writes descriptive summaries, but never diagnoses, prescribes or changes a dose, and every health answer defers to the treating physician. With no API key configured it degrades to fully manual and the app still works. Spanish-first and bilingual from day one, because that is the household it was built for. FastAPI and PostgreSQL on the server; React Native (Expo) on the client.
7 milestones, 6 complete0 AI diagnoses by designES / EN bilingual
FastAPI · SQLAlchemy 2.0 · PostgreSQL · Alembic · JWT auth
In development — App Store polish stage. A logging and reminder tool, not a medical device: it never diagnoses, prescribes or recommends. Private repository.
Engines Health Monitor
Predictive Maintenance · Production · NLP / LLMs
Fleet remaining-useful-life you can query in English — and download as a briefing.
An end-to-end predictive-maintenance service on NASA's C-MAPSS turbofan degradation data. XGBoost estimates remaining useful life per engine; SHAP attributes the prediction back to individual sensors so a technician can see which reading moved. FastAPI serves inference; a Streamlit dashboard is the fleet surface. On top of that, a LangGraph multi-agent layer answers diagnostic questions in plain English and writes a ReportLab PDF maintenance briefing. The whole stack is deployed on Google Cloud Run, so the demo is the production shape, not a local screenshot. Test RMSE 16.74 cycles, MAE 12.33, R² 0.825.
16.74 test RMSE (cycles)12.33 test MAE0.825 test R²15.02 val RMSE
XGBoost · SHAP · FastAPI · Streamlit · LangGraph
Personal Research Assistant
NLP / LLMs · Production
Chat with your PDFs. Every answer shows the snippet it came from.
A retrieval-augmented generation service: upload a PDF, chunk at 1000 tokens with 100-token overlap, embed with text-embedding-3-small, store in Chroma, retrieve through a MultiQueryRetriever, then answer with GPT-4o grounded only in those chunks — sources always visible. Built and tested against the HRV thesis itself, indexed to 149 segments covering features such as RMSSD and pNN50. FastAPI behind a Streamlit client, so the same pattern can sit in front of any document set, not only this one.
149 indexed segments1000/100 chunk / overlap100% answers cite source
FastAPI · LangChain · ChromaDB · GPT-4o · text-embedding-3-small
ADHD & Sex Prediction
Healthcare AI
Screening support with fairness and explanations, not a black box.
1,213 participants. Mutual-information feature selection, KernelPCA and stratified splits feed a Random Forest for ADHD and Logistic Regression for sex, with the sex model tuned toward female recall. LIME and SHAP carry the explanations so a finding can be inspected, not only scored. Fairness is reported by sex on purpose. ADHD accuracy 82%, recall 0.88–0.90, AUC 0.866; sex recall 0.71 male / 0.48 female — the gap is on the page, not in an appendix.
82% ADHD accuracy0.88–0.90 ADHD recall0.866 ADHD AUC0.71 / 0.48 sex recall M / F
scikit-learn · XGBoost · Keras · SHAP · LIME
Academic research. Not a clinical device.
Cardiovascular Risk Classification
Healthcare AI
Leak-free clinical ML — and proof that a pooled AUC can lie.
920 patients across four sites. Sentinel zeros in cholesterol and resting blood pressure are mapped to NaN, and the ColumnTransformer lives inside cross-validation so nothing leaks from the held-out fold. HistGradientBoosting is selected on recall then F1, explained with SHAP and LIME. Leave-One-Site-Out then exposes the transportability gap the pooled score hides: test ROC-AUC 0.90, recall 87.3%, F1 0.85 — LOSO mean AUC 0.79, worst site ≈ 0.70. Top drivers: asymptomatic chest pain, oldpeak, exercise-induced angina.
0.90 test ROC-AUC87.3% recall0.85 F10.79 LOSO mean AUC
scikit-learn · HistGradientBoosting · SHAP · LIME · ColumnTransformer
Educational. Not a validated clinical instrument. Worst-site AUC ≈ 0.70.
Motor Performance Monitoring
Predictive Maintenance
Interpretable machine health from vibration and temperature — and a first place for it.
An industrial preventive-maintenance system built around live vibration and temperature. It profiles two very different machines — a Sumitomo three-phase induction motor baselined across five speeds (50/60/75/90/100%) and a Haas Mini Mill baselined from combined operational trajectories — then applies machine-specific adaptive thresholds, because a milling machine is legitimately noisier than a motor. A threaded serial reader with auto-reconnection feeds the sensors into 0–100 health scores in a Streamlit dashboard, with replay mode for demos without hardware and live packet/FPS diagnostics. Deliberately rule-based rather than ML: every alarm traces back to the statistic that fired it, which is what a maintenance team can act on and audit. First place at Tec Expo Ingenierías; recognized by Rockwell Automation for contribution to an integration project using their technology (Nuevo León, December 2025).
1st Tec Expo Ingenierías2 machine types profiled5 speed baselines0–100 health score
Python · Streamlit · Signal processing · Serial / IoT · Statistical thresholds
First place at Tec Expo Ingenierías, the Tecnológico de Monterrey engineering project competition. Recognized by Rockwell Automation for contribution to an integration project using their technology (Nuevo León, Dec 2025).
HRV Emotion Recognition
Healthcare AI
From raw PPG to emotion models under leave-one-participant-out.
MSc thesis at the University of Essex. Raw photoplethysmography is Butterworth-filtered; inter-beat intervals are extracted; time-domain, frequency-domain (Welch PSD) and non-linear HRV features — including Poincaré descriptors — are computed and normalized against a per-subject calibration baseline. Mutual-information selection, nested grid search and leave-one-participant-out / leave-one-group-out validation compare Random Forest, XGBoost and SVR for continuous arousal. Best result: Pearson r = 0.5975, RMSE 0.154 ± 0.032, across 62 participants. LOPO is the point: the model has to work on a person it was not trained on.
0.5975 Pearson r (arousal)0.154 RMSE ± 0.03262 participants
Python · NeuroKit2 · SciPy · scikit-learn · Signal processing
MSc Artificial Intelligence thesis, University of Essex — Random Forest, LOPO validated. Awarded Distinction.
WellHave
Healthcare AI · NLP / LLMs
Daily pulse in; a risk level and a recovery plan out.
A three-tier classifier sorts a daily pulse into low, moderate or high burnout risk, then GPT-4o through LangChain turns that signal into a concrete recovery plan — not a pep talk. FastAPI backend, React Native (Expo) client, Supabase for state. The model decides the tier; the language model is only allowed to coach inside it.
3 burnout tiersMobile Expo client
FastAPI · GPT-4o · LangChain · React Native · Expo
Advanced Contact Segmentation
Production
Clustering that turns a contact list into an engagement strategy.
Proof of concept for Interius / APREU: unsupervised segments by engagement, geography × engagement, and channel activity, packaged as a Streamlit app, containerized and deployed to Google Cloud Run so stakeholders could open it themselves. The same 'notebook to Cloud Run' shape as the later RUL dashboard.
scikit-learn · Clustering · Streamlit · Docker · Cloud Run
California House Price Estimator
Forecasting · Production
Notebook to endpoint, without the gap in between.
Exploratory analysis leads into a RandomForest pipeline, served behind a FastAPI /predict route with a responsive HTML frontend on top. The exercise is the missing middle of most coursework: the trained model is an API, not a cell at the bottom of a notebook.
scikit-learn · RandomForest · FastAPI · HTML
Rossmann Sales Forecasting
Forecasting
1,115 stores, a six-week horizon, one feed-forward net.
A 128-64-32 MLP with 0.3 dropout, Adam and early stopping forecasts daily sales across a six-week window for 1,115 Rossmann stores, trained on engineered calendar, promotion and store features. RMSE ~0.06 on this project's engineered feature scale (the official Kaggle metric is RMSPE — noted so the number is not over-read).
1,115 stores~0.06 RMSE on engineered features
TensorFlow · Keras · Feature engineering · Pandas
RMSE reported on this project's engineered feature scale; the official Kaggle metric is RMSPE.
Text Emotion Recognition
NLP / LLMs
Six emotions across 416k tweets.
Bag-of-words with stemming feeds an XGBoost classifier, validated with stratified K-fold across six emotion classes on roughly 416,000 tweets. 87.5% accuracy. A classical NLP baseline — useful as the 'before transformers' control, and as proof the same boosting family used on sensors also holds on text.
87.5% accuracy~416k tweets
XGBoost · NLTK · Bag-of-Words · scikit-learn
Hybrid Movie Recommender
NLP / LLMs
Collaborative and content-based, then both at once.
SVD collaborative filtering, TF-IDF content similarity, and a hybrid blend of the two, wrapped in a Flask interface. SVD RMSE 0.477 against a 3.624 baseline — the collaborative half is doing the work; the content half covers the cold-start the SVD cannot.
0.477 SVD RMSE3.624 baseline RMSE
Surprise / SVD · TF-IDF · Flask · Pandas
Politrauma
Healthcare AI · Production
Polytrauma workflows, walked step by step.
An interactive clinical-education web tool that guides students through polytrauma assessment and trauma-care decision paths. Built for teaching, not for the bedside: TypeScript on Vercel, open so a cohort can walk the same case. Explicitly not a medical device.
TypeScript · HTML · Vercel
Educational. Not a medical device.
TurtleBot Navigation
Robotics
PID and fuzzy logic deciding in real time.
Edge following and obstacle avoidance on a TurtleBot, with fuzzy rules and PID control acting on live sensor readings. Intelligent Systems and Robotics coursework at Essex — the control-theory half of the mechatronics degree, running on a robot rather than in Simulink.
Python · ROS · Fuzzy logic · PID
Rocket Landing Neural Network
Robotics
A neural network written from scratch, no framework.
Forward pass, backpropagation, Xavier initialization and input normalization implemented by hand in NumPy, then flown against simulated landings. The point of the exercise is that the later TensorFlow work is not magic — the update rule was written out.
Python · NumPy · Backpropagation
