Agent pipeline architecture · AI infrastructure · ROCm multi-GPU
Paris, France · Self-taught
I build systems that find things people missed, and I make infrastructure run where the tooling says it shouldn't. Two examples. An agent pipeline that isolated 95 defects across 81 open-source repositories in 101 hours, and got its fixes merged into V8, LLVM, React and Tokio. And AMD GPUs doing tensor-parallel diffusion inference that the ecosystem assumes needs NVIDIA.
I don't write the patches by hand. I design the system that produces them, and the gates that decide what is allowed to leave it. I own the machine it all runs on, so every number here is measured, not estimated.
I built a system that hunts defects in large codebases, proves them, and prepares the fix for review. Two rules decide what is allowed to leave it: nothing ships without a reproduction that fails before the patch and passes after, and nothing ships until the target project's own contribution rules are satisfied. Everything else is engineering around those two.
What that produces, measured:
| 190 agent runs over 101 hours | 10–14 September 2026 |
| 95 defects isolated, proven and written up | across 81 repositories |
| 37 merged, 7 closed without merging | in 27 projects I don't maintain |
| $482 total — $16 per merged fix | failed runs included; half of them find nothing |
Seven rejections against thirty-seven merges is the number I care about, because that ratio is what the proof requirement is for. Automated reviewers land the same way — Copilot's reviewer returned approval recommended on the WSL parser fix, CodeRabbit called the cuDF one suitable for merge.
The hardest review it has passed is the JavaScript engine v8/v8: two CLs through Gerrit, CLA and committer review, into the ECMA-262 implementation behind Chrome and Node.js. My name is in the AUTHORS file.
The fixes I would point to first:
| tokio · Rust · 33k★ | copy_buf never touched the coop budget, so a task copying in a loop never yielded back to the runtime. The cooperativeness test never finishes without the fix. |
| React · 250k★ | Flight and DevTools only stripped part of the async prefix V8 puts on stack frame names, so server-component frames came out malformed. |
| celery · Python · 29k★ | AzureBlockBlobBackend.as_uri leaked the SAS token and the account key into logs and error messages. |
| swc · Rust · 34k★ | The TypeScript parser ignored a line break after declare, accepting code tsc rejects. Needed the tsc-references snapshots regenerated too. |
| MAME · C++ · 10k★ | Several WE32100 mnemonics were wrong in the disassembler. Proven against the AT&T instruction set manual, without a full MAME build. |
| LLVM · 40k★ | cmake_format.py in libc wrote CRLF on Windows, so formatting a file changed every line of it. |
| three.js · 116k★ | NURBSCurve never overrode copy(), so clone() returned a curve with no degree, knots or control points — a different shape from the original. |
Both counts above are checkable: merged · closed without merging.
The rest
Systems and infrastructure — kubescape (Go, /dev/stdout and /dev/null in diff --output) · authentik (translate before interpolating in SMS blueprints) · nginx-ui (bracket IPv6 hosts for the default gRPC port) · termux-packages · uWebSockets.js · sea-orm (Rust, multi-hop left_join_linked aliasing) · OpenSandbox
ML and AMD — pytorch/ao (propagate non_blocking in TorchAOBaseTensor._to_copy) · pytorch/ao (invalid escape sequences, W605 enabled so they stay fixed) · llama.rn · agent-framework (reset $LASTEXITCODE per command in persistent PowerShell sessions)
Editors, UI and desktop — lexical (TextNode setters read the latest state) · AFFiNE (accumulate overlapping doc priority requests) · MarkText (exported links from folders named with #, ? or %) · Sparkle (Objective-C, release-notes content length from the appcast) · mango (keycode-only modifiers in parse_mod)
Emulation and low level — MAME (fs_prodos: zero master index entries in tree files are sparse) · xiaozhi-esp32 and #2257 (C++ on ESP32: don't abort on short theme colors, free cJSON strings)
Tools — davinci-resolve-mcp (six merges on LUT install and media analysis) · pyvideotrans · QuantDinger (infer exchange precision from the Decimal exponent) · diagram-design
Orodruin · live at orodruin.dev — Open-source intelligence platform, a self-hostable alternative to Palantir Gotham. 30+ live public sources (GDELT, NASA FIRMS, USGS, AIS vessels, ADS-B aircraft, submarine cables, power grid, cyber threats, 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 · demo at excalibur.nextgen-labs.net — Active-defense deception middleware. Serves booby-trapped data to attackers, then traces and geolocates them through canary tokens. One-line integration · Rust SDK · ML scoring (PASS/CANARY/DECEIVE) · forensic tracer with Merkle log and STIX export · real-time SOC dashboard. AGPL-3.0.
flux-amd-rocm — FLUX.1-dev at NVIDIA parity on AMD RDNA3. Megatron-style tensor parallelism across 4 GPUs, 51 s per 1024² image at 11 GB per GPU. Int8 quantization with async group offloading brings it down to a single RX 7800 XT: 80 s, 12.5 GB VRAM.
diffusers-rocm-parallel — Multi-GPU inference stack for AMD. Tensor-parallel FLUX on gfx1101, ring attention LSE shape fix, Ulysses context parallel.
openclaw — Autonomous security research pipeline. Multi-agent orchestration on Qwen3 80B + 14B, fully local, recon → scan → CVSS → report.
Client work — systems I was paid to build, and still operate
Self-hosted email infrastructure — SMTP on my own IPs with full SPF, DKIM and DMARC alignment and automated warm-up, scoring 10/10 on Mail-Tester from the first send. Deliverability is an infrastructure problem, not a template problem. In production.
Cross-asset atomic swap — proof of concept on Daml/Canton for a client: the contract model, a 7/7 test suite, a web interface and bilingual documentation. Delivered.
Production hosting I run myself — client sites and apps on my own VPS fleet, provisioned and maintained end to end (nginx, pm2, certbot, PostgreSQL, Supabase). Among them rim-phone.com, a bilingual FR/AR storefront with full RTL and cash-on-delivery, and a booking system still in daily use.
Python · Node.js · Rust · Next.js · React · FastAPI · PostgreSQL · Daml/Canton · Docker
CAMUS Theory — Graft-based temporal cognition in frozen LLMs. A TemporalAdapter under 0.6% of parameters, grafted at mid-depth through a forward pre-hook, decodes log-time with R² ≈ 0.9 from 1B parameters. The subspace is roughly 5-dimensional and holds across model sizes. Validated on TinyLlama-1.1B and Qwen2.5-14B in under 30 minutes, for $0.83.
Published independently on Zenodo.
2022 – now · Freelance AI infrastructure, security research, independent publication.
2020 – 2022 · On-site GPU infrastructure engineer, 300+ GPU production facility, Kyiv, Ukraine. Hardware deployment, network architecture, 24/7 uptime under real production constraints.
2019 · Founded and shipped a repair-management SaaS solo — 350+ pages, logistics, billing, payments. Closed by Covid.
CPU 2× Intel Xeon E5-2698 v4 — 40 cores / 80 threads
RAM 512 GB ECC
GPU 6× AMD RX 7800 XT (gfx1101) — 96 GB VRAM
NVMe Multi-drive array
Net 10 GbE, self-hosted services
OS Ubuntu · ROCm 7.2.2 — custom builds: rocWMMA, FA_ALL_QUANTS, HIP_GRAPHS
Qwen3-Coder-Next 80B runs on it at 42 t/s with a 262K context. Everything above was built and measured here.
Inference PyTorch · diffusers · torchao · llama.cpp · vLLM 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 · Docker
Open to research collaborations, freelance infrastructure work, and projects that shouldn't exist yet.