Software that keeps working, and models that know when to stop.

Sameer Najm co-authored A Component-Isolated Analysis that confirmed the absence of an incremental signal of the Triglyceride–Glucose Index in CVD mortality. The authors received a peer-review acceptance notification from Natural Resources for Human Health; the journal now has an article page online.

The study found no added predictive signal from TyG in the analyzed models once fasting triglycerides and glucose were already included. It also reported limited performance when the model was tested on data from a different time period. These are research findings, not a clinical prediction service.

Read the article on the journal site →

Judges whether a clinical claim is entailed, contradicted or neutral with respect to given evidence, and abstains below a calibrated threshold. Intended as a verification layer beneath other systems, not as an answer generator.

Archived preprint: doi.org/10.5281/zenodo.20683289 · Not peer reviewed · Submitted to PhysioNet for credentialed model release, June 2026, awaiting editor assignment.

An original transformer architecture trained from scratch in-house. It proposes in latent space and is verified by a separate judge model, so the component that generates an answer is never the component that approves it.

Test counts measure software correctness, not model accuracy. The two are reported separately throughout this page.

A research interface combining a drug knowledge base with contraindication checking, laboratory reference data, emergency protocol templates and clinical calculators, over an explicit evidence chain.

The 96.6% figure is an internal evaluation on an internally constructed case set. It is not an external validation, not a clinical trial, and not evidence of clinical safety.

Twelve-lead electrocardiogram classification across twelve diagnostic classes, trained on the public PhysioNet PTB-XL dataset.

Benchmark performance on a public research dataset. Manuscript in preparation; not submitted, not peer reviewed.

Chest X-ray classification across fourteen findings, fine-tuned from a DenseNet121 base pretrained on the public CheXpert dataset.

Internal evaluation on held-out public data. High AUROC on a research dataset does not transfer automatically to clinical populations.

Defensive-only configuration auditing and hardening behind a default-deny human approval gate, with a written charter and a request classifier that redirects offensive requests to defensive guidance.

All figures are internal evaluations by the authors on internally constructed case sets. No independent red-team assessment has been performed.

Distilling a robust natural-language-inference student model, evaluated on public benchmarks including SNLI, MNLI, ANLI and XNLI.

Public benchmark results. Manuscript in preparation; not submitted.

An Arabic-language assistant for environmental standards, covering EPA, WHO and Iraqi IQS-417 pollutant limits, fine-tuned from an open 1.1B base and running in 2.2 GB of memory.

Accuracy figures come from an internal question set written by the same team that built the model. Treat as internal evidence only.

A default-deny gate for code changes. Evidence is bound to an exact repository state and patch, and each report declares the precise limits of what it inspected, so a verdict cannot quietly widen its own scope. The model may propose; only the deterministic verifiers decide.