Decision Support Systems · AI & Data Science

AI that makes people better decision-makers — not AI that decides for them.

I build systems that sit next to human judgment, not in front of it: transparent models, honest uncertainty, and interfaces that explain themselves. At the center is a mission — augmenting the people who care for patients — with the same principle proven across workforce, credit, aerospace, and customer-retention decisions.

See the work Read the philosophy
Philosophy

On Human–AI Relationship

Every project on this site is built on the same working belief: the goal of applied AI is not to replace a person's judgment but to make that judgment sharper, faster, and better informed. A model that quietly makes the call is a liability. A model that shows its reasoning, flags what it doesn't know, and hands a clear decision back to a human — that's a tool worth trusting.

01

Augment, don't replace

Systems are designed as decision support — surfacing patterns, risks, and probabilities for a human to weigh, rather than issuing verdicts a person is expected to accept on faith.

02

Transparency over accuracy theater

A model's headline metric is never the whole story. Every project here documents what the data can't tell you — missingness, top-coding, sampling bias — alongside what the model gets right.

03

Built for the person who has to act

A dashboard nobody reads and a model nobody trusts are the same failure. Every system is designed around the actual decision-maker: what they need to see, in language they already use.

The Pento‑Helix

Underneath those three principles sits a broader frame I work from: the Pento‑Helix. Pento, for five — a structure meant to stay portable, flexible, and adaptable across contexts, not fixed to one project. Helix, because it behaves like DNA: one constant structure that expresses itself differently everywhere it's deployed, the way RNA carries out what DNA encodes.

01Human–AI collaboration and synchronization
02Human–AI harmony
03Each side strengthening the other's skills and expertise, while understanding the other's limitations — mutual respect
04Shared social responsibility for what gets built
05Progressive, together — the constant thread across every project, and what this site's mark is built around: a dot held steady at the center — the Tao — with a spiral of motion, a Milky Way, turning around it.

The first four pillars above are how the Pento‑Helix takes shape in the Human–AI Augment Systems work collected on this site. Other projects will express their own version of the same four; the fifth never changes.

Selected Work

Seven decision-support systems — human-centered care at the heart

Every system here rests on one belief: AI should make people better decision-makers, not replace them. That belief has a home — the Nurse Augmenting System, human-centered patient care and the core of the AEGIS programme. The rest prove the same principle holds across domains: two are deployed as live, working apps, and all carry honest, full write-ups of what each model can and can't tell you.

The Core

Nurse Augmenting System — the heart of the work

The Pento‑Helix vertex: augmenting the people who care for patients, never replacing their judgment. Two studies in human-centered healthcare, built under the AEGIS research programme in the tradition of Prof. Dhanjoo N. Ghista.

Healthcare · Core Project

Nurse Augmenting System for Biomedical Intelligence and Personalized Care Augmentation

Eight independently trained models — spanning linear, tree, kernel, boosting, and ensemble methods — converged on the same five predictors of 30-day heart-failure readmission. A hand-engineered clinical score (HCAS), not raw accuracy, was the point.

8 / 8
models agree on
the top 5 predictors
5
predictors in the
HCAS clinical score
30-day
readmission
window
View case study
Healthcare · Core Project

Medication Adherence & Healthcare Claims Analysis

A star-schema claims model and a five-model adherence benchmark on 24,084 diabetes and hypertension patients — surfacing that non-adherent patients cost 2.5× more, and including an honest report of the one model, out of five, that quietly failed.

81.98%
champion accuracy
(Gradient Boosting)
2.5×
claims cost of a
non-adherent patient
24,084
patients
modeled
View case study
Deployed & Working

Systems that ship — and one built to raise the alarm

Two live, human-in-the-loop apps you can use right now. AEGIS Orbital Watch is the outreach signal — a small voice on satellite-collision risk in a very large ocean. Predictive Maintenance is the working proof the pipeline runs end to end.

Featured · Full Case Study · Deployed

AEGIS Orbital Watch: Advance-Warning System for Satellite Collision Risk

Built on ESA's Kelvins Collision Avoidance dataset (13,154 events) — an eight-model benchmark across four increasingly rigorous passes, culminating in a deployed, human-in-the-loop app that scores live conjunction data and lets an operator record their own call alongside the model's. Built to raise awareness of a growing orbital problem.

Python / pandas XGBoost FastAPI Human-in-the-loop UI
View case study
0.9814
Test ROC-AUC (XGBoost)
80.8%
Recall @ operating threshold
13,154
Conjunction events modeled
Featured · Full Case Study · Deployed

AI-Powered Predictive Maintenance Decision Support System

An eight-model benchmark on the AI4I 2020 dataset (10,000 machines, a 3.4% failure rate), run across two controlled experiments — where two engineered features lifted every model and decided the champion. The five built-in failure-mode flags were removed first to prevent target leakage, and the operating threshold was tuned for advance warning, not headline accuracy. A Flask app puts the model behind a technician-facing form with a red / amber / green verdict.

Python / pandas scikit-learn Random Forest Flask Feature engineering
View case study
0.9766
Test ROC-AUC (Random Forest)
80.9%
Recall @ advance-warning threshold
10,000
Machines modeled
The Method, Across Domains

The same discipline in workforce, finance, and marketing

The reusable AEGIS framework, applied and validated on public datasets in three more domains — same rigor, same honesty about what each model can and can't do.

Featured · Closed · Framework Validated

AI-Powered Workforce Income Analytics & Decision Support System

A unified master dataset (48,790 U.S. Census records) feeding a Power BI analytics layer and a reusable AI framework that benchmarked five ML algorithms across two feature-engineering experiments — closed once the champion model reached the published literature ceiling for this benchmark.

Python / pandas scikit-learn XGBoost CatBoost Power BI / DAX
View case study
92.79%
Champion ROC-AUC (XGBoost)
87.96%
Champion accuracy (XGBoost)
48,790
Individuals modeled
Finance

AI-Powered Loan Default Risk Decision Support System

A 12-model benchmark on 29.9 million LendingClub records — where an obvious leak was caught and removed, and a suspiciously perfect score that remained was investigated and explained rather than just reported.

View case study
Marketing

AI-Powered Marketing Decision Support System for Customer Retention

Nine models were benchmarked on IBM's Telco Churn dataset — and the highest-accuracy one wasn't picked. The production model was chosen for matching a real business threshold instead.

View case study
Coming soon · In the pipeline

Predictive Remaining-Useful-Life (RUL) for NASA turbofan engines

The regression companion to Predictive Maintenance — not "will it fail?" but "how many cycles of life are left?" A reusable pipeline benchmarking eight regressors across NASA's four C-MAPSS turbofan datasets, with an asymmetric scoring rule that penalizes dangerous late predictions harder than early ones. Full case study in progress.