Dr. Sithira Ambepitiya

Synthetic images, real diagnosis

Peer-reviewed, Scientific Reports 2024. Ten real clinical photographs were enough. Diffusion models turned them into a training set for HPV-related penile disease; the classifier was measured against 210 real clinical images it had never seen: 91.4% sensitivity, 95.5% precision. Data-efficient diagnostic modeling without a large labelled patient cohort. Second author.

https://www.nature.com/articles/s41598-024-77565-6

Mobile diagnostic platform

Peer-reviewed, Mayo Clinic Proceedings: Digital Health 2024. Five penile diseases told apart from a phone photo. Mean per-class F1 of 0.944 over a held-out set of 239 images, specificity above 0.90 for every disease. Second author. Commercial context: HeHealth.

https://doi.org/10.1016/j.mcpdig.2024.04.006

Mpox detection advanced

Preprint, arXiv 2024. Training images generated from 120 clinically validated guides; classifier scored 97% accuracy on 300 real patient photographs across mpox, normal, and other skin conditions. Internal test set — not externally validated.

https://arxiv.org/abs/2407.17762

SynthVision

Preprint, arXiv 2024. Ten guide images in, a working HPV classifier out: 97% accuracy and 0.993 AUC against 140 real clinical photographs. A physician discarded roughly a fifth of generated images as clinically implausible before training — clinician in the loop.

https://arxiv.org/abs/2402.02826

Otiz — STI counseling chatbot

Preprint, arXiv 2024. LLM-based chatbot for counseling patients with sexually transmitted infections and genital diseases. Evaluated by venereologists on diagnostic accuracy, relevance, correctness, comprehensibility, and empathy.

https://arxiv.org/abs/2412.12166

Crimson Dawn

Agentic engineering harness: request to pull request. Every task carries a scope contract — paths it owns, paths it must never touch — enforced while the edit is happening. Risk-tiered models, bounded loops, merges decided by tests, independent review, in-scope diffs, and no sensitive paths.

Stasis

Compute runtime that freeze-thaws agents at the model-call boundary so waiting workers do not pin RAM. Snapshot, complete the provider call, restore.

TruGhostOS

Agent-native Linux distribution. Workers confined by cgroups, namespaces, and seccomp; OS mutation gated behind snapshots with rollback.

Osiris and Sentinel

Persistent agent fleet. Osiris is a root agent on the Mac that survives reboots; Sentinel routes work in from Discord. https://github.com/CrimsonSithria

T3MP3ST

Multi-agent offensive-security fork of Pliny the Elder’s T3MP3ST, extended for headless workers and tested against a home SOC lab. Authorized targets only. https://github.com/elder-plinius/T3MP3ST

Agent Readiness

Audits a codebase against 82 criteria across nine categories and returns an evidence-backed report with a prioritised fix list before coding agents are adopted.