AI Infrastructure · Multi-GPU ROCm · Independent Research · Offensive Security
Full Stack Development
Paris, France · Self-taught · No degree · Full stack from silicon to inference.
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.
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
SaaS platforms, mobile apps, full-stack web. Recent deliveries:
- Email marketing platform — self-hosted, SPF/DKIM/DMARC, warm-up automation, 10/10 deliverability on first test
- Yoga studio app — React/Vite, Supabase auth, booking system, deployed in production
- Hyperlocal marketplace — mobile app, real-time geolocation, neighbor-to-neighbor listings
- OSINT platform — osint-platform — open-source Palantir alternative, 6-tier data ingestion, entity graph, real-time analysis
Stack: Python · Node.js · Rust · Next.js · React · FastAPI · PostgreSQL · Supabase · 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.