Selected work / 01–04

Questions into systems.

Research and engineering across visual data, biological simulation, and applied AI. Each project has a different scope; the evidence and limitations are part of the work.

01 / Master’s thesis · DFKICHIRA 2026 · Accepted

DiversityLens

Auditing the demographic structure of visual datasets.

Watch the project walkthrough

The question: What demographic patterns are hidden inside the datasets used to train and evaluate visual AI?
My contribution: A Python pipeline spanning discovery and sampling, face detection, attribute inference, aggregation, and reporting.
Research outcome: Audits across 1.6M+ frames and 10+ benchmark datasets. Developed as master’s thesis research at DFKI, with a paper accepted at CHIRA 2026.
Interpretation: Inferred demographic attributes depend on the models used. Dataset coverage and model uncertainty matter when interpreting the audit.

Read the case study ↗View code ↗

02 / Independent research softwareExploratory · In development

Cell Engine

Connecting biochemical state, stochastic processes, and a visible cell.

Watch the project walkthrough

The question: How can interacting processes be represented in an inspectable model of a hepatocyte?
Engineering focus: Simulation architecture, traceable parameters, automated checks, and an interactive interface for examining the model.
Try it: Open the demo, compare the control with BSEP loss, then inspect the engine response. Change the hepatic zone or nutritional state to explore the available snapshots.
What it demonstrates: A research software prototype. The browser combines precomputed engine snapshots with local visual animation; motion is not a live Python simulation. Biological validation remains incomplete.

270+ tests are reported for the project. Software tests alone do not establish biological validity.

Launch interactive demo ↗View code ↗

03 / Retrieval & source evidenceStatic demonstration

AI Evidence Assistant

A document question-answering project with upload, search, and cited passages. The portfolio demo uses bundled sample evidence and scripted responses to show the interface.

No live model or document processing runs in this demo. Sample scores are illustrative.

Open sample demo ↗Code ↗

04 / Training → tensor exportWork in progress

Edge AI

A compact 7–12–8–1 network trained with manual tensor operations, with learned weights exported for a C++ loader. Full C++ inference and int8 quantization remain future work.

The displayed loss and accuracy curves are illustrative, not recorded training results.

Explore the pipeline ↗Code ↗

Research & AI/ML engineering

Have a related problem?

I’m open to engineering roles, research collaborations, and PhD opportunities.

Get in touch ↗