• Education: High school student focused on AI development.
• Focus: Deep learning, computer vision architectures, and multimodal models (CLIP).
• Achievements: AI Challenge — Participant in the Main Stage of the international youth AI competition.

SysSpectogram is a Linux / VPS defense utility that treats host telemetry as images.

It samples CPU, memory, network, disk, and GPU metrics every second, builds 60-second windows into spectrogram-like matrices, and scores them with a CNN (computer vision) plus Isolation Forest (classical ML) ensemble — useful for behavioral anomalies such as CPU / GPU crypto-miners, resource abuse, and unusual host load. Around that core it adds perimeter / egress watching, OSINT on the attacker IP (nmap, DNS, certificate logs), and a Telegram / web control plane for alerts and SOAR-lite responses (ban, recon, kill — with dry-run). In v1.0 it also ships a Tauri 2 + Svelte Control Center desktop app, a polkit-gated root helper, a Rust integrity agent with Aya eBPF probes (execve, openat, module load, kill, setuid), and an incident engine with ATT&CK mapping and HTML reports.

Gavel AI is an Apache-2.0 computer-vision prototype that looks for signs of AI image generation — not a courtroom “this is fake” claim, but a layered signal you can inspect.

L0 is a frozen OpenCLIP ViT-L/14 backbone with a calibrated linear probe (generated / uncertain / real). Around it sit local layers: anatomy (MediaPipe face / pose / hands map + strict 2D hand warnings), artifacts (flat-region residual kurtosis as a soft cue), geometry (LSD + vanishing-point consistency), and a physics stub. Fusion A combines them: L0 or a hard layer vote can raise a generation signal; soft artifacts cannot overturn a strong L0=real; a clean face-free hand close-up can veto a known L0 false positive on stock gym hands. Wide diptychs are scored on left/right tiles. Ships a Gradio UI, typed EvidenceProtocol JSON, OpenCV overlays, and training / eval tooling for the probe.

Languages

Machine Learning & CV

Tools, UI & OS