Research portfolio

Trustworthy ML for messy, human worlds

I build machine-learning systems for high-stakes, real-world settings: crops, clinics, and faces. My work asks what happens after accuracy — when a model meets unfamiliar data, needs to explain a decision, or must earn a person's trust.

Explore the work

Discuss research

Featured case study

TrustinAgrAI

A trustworthy decision-support system for grape-leaf health assessment.

Active researchComputer vision · agriculture

A strong classifier is not automatically a trustworthy one.

The first TrustinXAI application classifies healthy grape leaves, Downy Mildew, and Powdery Mildew from field-captured images. A fourth disease is deliberately held out to test whether the model can recognise when it does not know.

The question

Can a useful plant-disease model remain calibrated, show where it is looking, and flag unfamiliar symptoms instead of making a confident wrong prediction?

The approach

EfficientNet-B0 provides the baseline. Temperature scaling, Grad-CAM, and Mahalanobis distance add calibration, visual explanation, and out-of-distribution detection.

What I learned

The baseline reached high test accuracy, yet softmax still gave unseen disease images a mean confidence of 88.93%. Accuracy was the beginning of the trust problem, not the end.

Now

Phase 1 training is complete. The current work evaluates calibration, explanation quality, and whether the OOD detector separates unfamiliar disease from known classes.

PyTorchEfficientNet-B0Grad-CAMCalibrationOOD detection

Run artifacts · Phase 1

Inspect the baseline, not just the headline.

These figures and numbers come from the saved training run. Switch views to inspect model behaviour by class and see why the held-out disease became the next research question.

Confusion matrix showing 142 correct Downy Mildew predictions, 188 correct Healthy Leaves predictions, and 59 correct Powdery Mildew predictions — **391 of 394** test images landed on the diagonal. The three misses are visible rather than folded into the accuracy score.

Recorded evaluation

98.73%: Accuracy
0.9863: Macro F1
394: Test images
3: Known classes

Dataset split: 1,838 train · 394 validation · 394 test · 100 unseen-disease probes.

Related work

Research prototypes

Different domains, one recurring question: how do we build systems people can inspect, question, and safely use?

TriageAI

Prototype

Offline triage decision-support for hospital nurses. It turns symptoms, basic vitals, and history into a structured recommendation while keeping patient data on the device.

Question: Private clinical AI
Stack: Python · Gemma · Ollama
Role: End-to-end build

View repository Project details →

Realistic Filters

Stage A complete

A computer-vision pipeline for virtual makeup that preserves skin texture, facial geometry, identity marks, and skin tone by construction.

Question: Fair image enhancement
Evidence: 30 / 30 tests passing
Stack: OpenCV · MediaPipe

View repository Project details →

Research principles

What I measure beyond accuracy

Uncertainty

A model should know when an input falls outside what it learned and make that uncertainty visible.

Interpretability

Explanations should help a domain expert challenge the model, not merely decorate its prediction.

Human oversight

In consequential settings, the system supports a trained person and leaves the final decision with them.

Fairness by design

Identity, skin tone, privacy, and access are architectural constraints — not checks added at the end.

Open to collaboration

Research should survive contact with reality.

I am interested in trustworthy AI, computer vision, applied ML, and work that connects technical rigor with human needs.

Start a conversation