AI Researcher
I'm MD Mizanur Rahman. My research interests connect medical imaging, machine learning, and evidence-aware language systems. Explore the research and engineering demonstrations below to see how inputs, model outputs, and supporting evidence relate.
Medical imaging · Machine learning · RAG · Knowledge graphs
Schematic · data, model and analysis stages · synthetic, not clinical imagery
Inspect a stage
Select a stage to see what it does and what it is not allowed to do.
Featured
The lead demonstration is a synthetic workbench, not a published result. It shows how retrieved material is kept from becoming patient evidence: which stage decides, on what input, and what gets withheld.
Longitudinal reporting
An offline, inspectable state machine for interval-aware retrieval, patient-only candidate verification, source-aware admission, report realisation, one constrained rewrite and template fallback, running on authored synthetic cases.
Retrieval and graphs
A companion explorer that builds a typed graph over a small fictional corpus, retrieves passages through lexical and graph-neighbourhood paths, and renders answers that cite only identifiers that resolve.
Longitudinal reporting
A versioned evaluator for structured report content: canonicalisation, one-to-one exact matching, tuple precision and recall, and report-level false-content and omission states with explicit denominators.
Selected engineering
Retrieval and graphs
A synthetic benchmark that asks whether retrieval changes appropriately when the comparison interval changes, with explicit graded relevance labels and an ablation that removes temporal metadata.
Longitudinal reporting
Isolates the patient-only assessment boundary as a type: the evidence bundle has no field for retrieved content, and any attempt to pass it raises.
Medical vision
Modest, reproducible multi-label chest image baselines on a named MedMNIST task with a strict train/validate/test separation and a layer inspector that reports real shapes.
Interactive methods
Swap the earlier examination or the retrieved reference and see which claim is admitted, withheld or rejected, and why.
Open workbenchAsk a question of a small fictional corpus and inspect the typed path from every cited passage to its source.
Open explorerMove an admission threshold and watch precision, recall and omissions recompute from the visible fixture.
Open evaluatorResearch collection
Listed with publisher metadata only. Author order comes from the publisher record; no rankings or citation counts are shown.
Discover Artificial Intelligence, 6(1), article 833. DOI 10.1007/s44163-026-01492-w
Journal of Neonatal Surgery, 14(32S), 415–431. No DOI displayed in the publisher record.
International Journal on Science and Technology (IJSAT), 16(2), article 4206. DOI 10.71097/IJSAT.v16.i2.4206
International Journal on Science and Technology (IJSAT), 16(1), article 1751. DOI 10.71097/IJSAT.v16.i1.1751
Journal of Computer Science and Technology Studies, 3(2), 116–123. DOI 10.32996/jcsts.2021.3.2.9
Engineering practice
Preprocessing fitted on training data only; checkpoints chosen on validation; the test split touched once. Each repository carries a test that proves it.
Every run writes a manifest with the data version, split hash, configuration hash and seed. Smoke and demo runs are marked as such so they cannot be mistaken for experiments.
Adverse findings stay next to favourable ones. A number without a producing script does not appear.
An offline CPU path exists for every project. Restricted data is never bundled; adapters explain exactly what the user must obtain.
Background
I am pursuing master's studies in Data Science and Artificial Intelligence at Campbellsville University. My earlier work in IT and healthcare administration involved data records, reporting, and information handling. My current focus is on research and practical tools for machine learning, medical imaging, and retrieval-augmented systems.