# Denis Leclercq | On-premise infrastructure, cloud, and local AI # 0x0800 SRL | BCE: BE1021845609 | Liège, Belgium # Contact: contact@0x0800.com | https://0x0800.com ## Positioning 0x0800 SRL builds and operates infrastructure across three fields treated as a single discipline: on-premise systems, cloud and hybrid platforms, and local AI inference. A rack in a machine room, a cloud account, and an inference box sitting on a desk are three forms of the same work: size it right, automate what repeats, supervise what runs, and leave behind something the client's own teams can take over. Twenty years of infrastructure work back the delivery, from network support to technical direction. The market is companies in Belgium and France, served on site and remotely. ## What we do ### On-premise infrastructure Virtualised estates (Proxmox, VMware), networking and BGP routing, high availability, backup and supervision, maintenance in operational condition, redundant databases and front ends, inventory and operational documentation. ### Cloud and hybrid Scalable and resilient AWS and Azure architectures, migration of on-premise environments to the cloud, infrastructure as code with Terraform and Ansible, CI/CD pipelines with DevSecOps integration, continuous drift detection and compliance evidence. ### Local AI vLLM inference servers and quantised models, memory/thermal/context sizing, agents connected to internal sources, Hygur local edition deployment, team training and handover. Data stays on the client's hardware, which simplifies the RGPD legal basis (no transfers outside the EU) and makes EU AI Act obligations tractable when addressed early. ## Pricing Not published. Work is quoted per engagement after an on-site scoping day. Prices are communicated by nominative quote, valid 30 days. ## Published measurements Benchmarks are run on in-house hardware and released with their method, their commands verbatim, and their caveats. What was not verified is stated as such. - DeepSeek-V4-Flash-0731 on two DGX Spark: the complete recipe (4 August 2026) — https://0x0800.com/article-dsv4-2x-spark-recette.html A 43-layer, 256-expert model on two desktop boxes in tensor-parallel, roughly EUR 8,000 of hardware and 90 W peak for the pair. End-to-end setup: RoCE networking (declare both HCAs, forgetting the second costs 39% of collective bandwidth), venv and build, checkpoint transfer and header verification, launch, warm-up. Code quality judged by execution rather than opinion: 96/100 for the shipped configuration (Python 50/50, Go 46/50), zero truncations, failures always compile errors and never a wrong business test. Single-stream decode 78.55 tok/s on code and 47.16 on prose — prose is the slower case, which is counter-intuitive and reproducible. Context length has no measurable effect on decode step speed between 4k and 131k tokens (sparse attention reads a fixed 512-entry budget); what does cost is prefill, 87 s before the first token at 131k. Raising --max-num-seqs from 2 to 16 is the only tuning that produced a gain: 188.9 tok/s aggregated at sixteen streams with TTFT under 0.8 s, for 4.2% of KV pool. 1,793,656 tokens of KV cache, read from the startup log and not /metrics, which reports a figure twice too low. Prefix caching cuts prefill by a factor of 18 and survives across requests. No thermal or power throttling over 180 samples, 67 °C worst case, 41 W per box. RDMA latency 1.86 µs typical, latency-bound rather than bandwidth-bound. - DSv4 on 2x DGX Spark: assembly, measurements, tuning (4 August 2026) — https://0x0800.com/article-dsv4-2x-spark-rapport.html The reproducible companion report. Two software paths documented exactly: path A, the jasl/vllm fork built from source, which produced every figure here, and path B, the upstream repository's pinned runtime, verified only up to import and never measured — stated as such. Findings that go against the grain: the flashinfer_b12x MoE backend costs 19% of throughput and adds a deterministic ~31 s fixed cost per request beyond ~1850 tokens of context, while the unremarkable MARLIN kernel wins on both counts; nvfp4_ds_mla costs 6 points of code quality without returning any memory. A one-knob sweep over five arms produced no gain at all, with a ±2% error bar established by running the same configuration twice; two knobs did not do what their name suggested, and one warning from the engine never closes at any value. Twelve parameters discarded out of sixteen: the soberest configuration is the optimum. Single-stream throughput obeys a one-parameter law, throughput ~= 13.90 x (1 + tokens accepted per step), accurate to 1.5%, which caps it at 83.4 tok/s at k=5 — measured best point 84.03, so at the ceiling rather than near it. k=7 is measured and loses 11.5-17.1%: the step costs 12.8% more while accepted tokens per step stay identical at 4.30. torch.compile is not activatable on DSv4 NVIDIA and that is a design choice, not an oversight. Includes corrections of two earlier claims of ours, with the protocol defects that produced them. - Serving Laguna S 2.1 NVFP4 on a single DGX Spark (23 July 2026) — https://0x0800.com/article-laguna-s21-single-spark.html A 117.6B-parameter MoE (8.5B active) on one unified-memory box: memory hygiene, native 256K context, DFlash speculative decoding, 738,410 tokens of KV cache at 256K, 49–52 tok/s decode, 252 ms TTFT. - Running vLLM properly on a DGX Spark, GB10 (20 July 2026) — https://0x0800.com/article-dgx-spark-vllm-nvfp4.html From cloning vLLM to serving an NVFP4 quantised model: the three bugs that block the build and the first server on SM121 architecture, and how to fix them. ## Products ### Hygur (https://hygur.ai) Local-first personal AI. Data stays on the user's machine, the model comes to it. This is the product deployed on-premise in companies, available as a local edition and a multi-tenant sovereign cloud. ### CompliMetric (https://complimetric.com) Continuous cloud compliance SaaS. Scans cloud and IaC, detects drift, and produces NIS2 / ISO 27001 evidence without spreadsheets. Ships with a GitHub Action and an MCP server. Both products are edited and operated in-house by 0x0800 SRL. ## Method 1. Scoping: one day on site to review the existing setup, the intended use, and the data and compliance constraints. Output: a roadmap and a quote. 2. Build: installation, migration, or deployment, depending on the engagement — on-premise infrastructure, cloud environment, or an inference server connected to internal sources. A narrow scope first, production second. 3. Autonomy: documentation, team training, supervision. The client keeps control of what runs on their premises. No dependency created. ## Technical Expertise - On-premise / local AI: local LLM serving, inference hardware sizing, RAG on private data, LLMOps - Cloud Platforms: AWS, Azure, OVH, Scaleway - Infrastructure as Code: Terraform, Ansible - Containers & Orchestration: Kubernetes, Docker, ECS Fargate - CI/CD & DevSecOps: GitHub Actions, Jenkins, shift-left security - Monitoring: Prometheus, Grafana, CloudWatch, ELK - Compliance: RGPD, EU AI Act, NIS2, ISO 27001 - Networking: BGP, VLAN, Cisco, HAProxy, Nginx - Operating Systems: Linux (Debian, RHEL), Proxmox - Languages: Python, Go, TypeScript, Bash, C ## Professional Experience (founder, before and alongside 0x0800) These roles were held by Denis Leclercq as an employee or consultant, not as engagements contracted by 0x0800 SRL. - NRB | Senior Cloud Architect & DevOps Consultant (2024–present, Belgium) - FABERNOVEL | Lead Infrastructure & Cloud Engineer (2020–2024, Paris/Remote) - Vallendra | CTO & Co-Founder (2018–2020, Paris) - Addventa | Technical Consultant / DevSecOps (2017–2018, Paris) - Free Infrastructure | Technical Writer (2014–2017, France) - PCWorld France | Tech Journalist (2011–2013, France) - QOSMOS | Customer Support Engineer (2009–2011, Paris) ## Company Information - Company: 0x0800 SRL - Registration: BCE BE1021845609 - Location: Liège, Belgium - Website: https://0x0800.com - Email: contact@0x0800.com ## Social Links - LinkedIn: https://www.linkedin.com/in/denis-leclercq - GitHub: https://github.com/Mirrdhyn - Telegram: https://t.me/Mirrdhyn ## Languages - French: Native - English: Professional working proficiency ## Key Differentiators 1. One discipline across three fields: on-premise, cloud, and local AI handled by the same hands, so a hybrid estate is not split between three vendors. 2. Measurements published: benchmarks run on in-house hardware, released with commands, method, and caveats, and reproducible by anyone. 3. Data stays in-house: models run on the client's own hardware, no data leaves the company. 4. Editor, not reseller: Hygur and CompliMetric are designed and operated in-house. 5. Independence: teams are left autonomous after training; no dependency created.