I build systems that run at the edge of what's technically possible — locally, at scale, without compromise. No cloud dependency, no abstraction layers hiding the truth. Every component understood, every parameter owned.

From founding a SaaS startup at 26, to operating a 300+ GPU farm on-site in Ukraine, to deploying 80B LLMs on self-hosted ROCm infrastructure and publishing independent AI research — every step was built from scratch, under real constraints.

Orodruin — Open-source intelligence platform · live at orodruin.dev

A self-hostable Palantir Gotham alternative. 30+ live public sources (GDELT, NASA FIRMS, USGS, AIS vessels, ADS-B aircraft, submarine cables, power grid, cyber threats, weather, satellites) fused onto a 2D/3D map and an actor graph, with an AI analyst that queries every source and drives the interface. FR/EN/AR/RU · AGPL-3.0.

Excalibur — Active-defense deception middleware · live demo at excalibur.nextgen-labs.net

Serves booby-trapped fake data to attackers, then traces and geolocates them via canary tokens. One-line integration · Rust SDK · ML scoring (PASS/CANARY/DECEIVE) · forensic tracer (Merkle log, STIX export) · real-time SOC dashboard. AGPL-3.0.

flux-amd-rocm — FLUX.1-dev at parity with NVIDIA on AMD RDNA3

4-GPU Megatron-style tensor parallelism · 51 s/image @ 1024² · 11 GB/GPU. Int8 quantization + async group offloading on a single RX 7800 XT · 80 s · 12.5 GB VRAM.

diffusers-rocm-parallel — Multi-GPU inference stack for AMD

Tensor parallel FLUX on 5× RX 7800 XT (gfx1101) · ring attention LSE shape fix · Ulysses context parallel.

openclaw — Autonomous bug bounty pipeline

Multi-agent orchestration · Qwen3 80B + 14B · fully local · recon → scan → CVSS → report.

CAMUS Theory — Independent AI Research

Graft-based temporal cognition in frozen LLMs. TemporalAdapter (<0.6% params) grafted at mid-depth via forward pre-hook. R² ≈ 0.9 for log-time decoding from 1B parameters. ~5-dimensional subspace invariant across model sizes. Validated on TinyLlama-1.1B and Qwen2.5-14B in under 30 minutes for $0.83.

• 2020–2022 — On-site GPU infrastructure engineer, 300+ GPU production facility, Kyiv, Ukraine. End-to-end hardware deployment, network architecture, 24/7 uptime under real production constraints.
• 2019 — Founded and shipped a full SaaS repair management platform solo (350+ pages, logistics, billing, payments). Shut down by Covid.
• 2022–now — Freelance AI infra, security research, independent publications.

Compute 5× AMD RX 7800 XT (gfx1101) · 80 GB VRAM · ROCm 7.1
Custom builds: rocWMMA · FA_ALL_QUANTS · HIP_GRAPHS
Inference PyTorch · diffusers · torchao · llama.cpp · vLLM · 38 t/s @ 80B ctx 262K
ML Tensor parallelism · group offloading · int8/int4 · Triton kernels
Security nuclei · subfinder · katana · httpx · Burp Suite Pro · responsible disclosure
Systems Python · Rust · Node.js · Next.js · FastAPI · PostgreSQL · Supabase · Docker

Production SaaS, mobile apps and full-stack web — shipped end to end, most of it self-hosted. Selected deliveries:

• Email marketing platform — self-hosted SMTP with full SPF/DKIM/DMARC, automated IP warm-up, 10/10 on Mail-Tester on the first send. In production.
• E-commerce store (rim-phone.com) — Next.js + Prisma, bilingual FR/AR with full RTL, cash-on-delivery checkout. Live on HTTPS (VPS · nginx · pm2 · certbot).
• Atomic swap PoC — cross-asset atomic swap on Daml/Canton, 7/7 test suite, custom web UI, bilingual docs. Delivered to client.
• Yoga studio app — React/Vite, Supabase auth, full booking system. Deployed in production.
• Hyperlocal marketplace — mobile app, real-time geolocation, neighbor-to-neighbor listings.

Stack: Python · Node.js · Rust · Next.js · React · FastAPI · PostgreSQL · Supabase · Daml/Canton · Docker · Stripe · REST APIs

CPU 2× Intel Xeon E5-2698 v4 — 80 threads
RAM 512 GB ECC
GPU 5× AMD RX 7800 XT (gfx1101) — 80 GB VRAM total
NVMe Multi-drive storage array
OS Ubuntu · ROCm 7.1
Net 10 GbE local · self-hosted services

Open to research collabs, freelance infra missions, or projects that shouldn't exist yet.