I am Julien Compain, founder and lead engineer of NovaQuantiX — an independent AI engineering practice. I ship custom Model Context Protocol servers, fine-tuned open-weight models, and autonomous agent swarms for teams that need real artifacts instead of decks.
NovaQuantiX exists because I believe the value of an AI engagement is in the signed binary you can run on day 365, not in the slide deck delivered at the end of week 2. Every engagement leaves the client with source code, signing keys, evaluation baselines, and a runbook — zero lock-in by construction.
How I work
- Fixed-price phases. No hourly billing. Each phase has explicit acceptance criteria written into the SOW and a single milestone payment on delivery.
- Reproducibility first. Every build is signed with Ed25519, every log entry is hash-chained into a Merkle root, every evaluation is replayable byte-for-byte on day 365 with the same seeds.
- Memory-safe by default. Performance-critical paths in Rust. Low-level in C with sanitisers in CI. TypeScript for MCP servers and agent interfaces. Python only for training pipelines.
- Vendor-neutral. Open-weight models (DeepSeek V4, Kimi K2.6, GLM 5.1, Qwen 3.7, Gemma 4, Llama 4) over closed APIs whenever the quality bar allows. Standard protocols (MCP, OpenAI-compatible endpoints) over proprietary ones.
- Transparent communication. One-page proposals. One-line status updates. Weekly written report.
Engineering principles I commit to
- Never invent. If a benchmark, latency number, or cost figure is in a document, it is measured — not estimated.
- Never simulate. No placeholder code in production. No stub returning fake data unless explicitly labelled as a fixture inside a test file.
- Never hide failure. A failed eval, a missed SLO, a regression — reported in the same hour they are observed, with a root cause hypothesis attached.
- Always sign. Every release artifact is verifiable. No tag, no signature, no deploy.
Domains I focus on
- Model Context Protocol. Schema design, stateful tools, streaming I/O, OAuth/mTLS, production hardening.
- Open-weight fine-tuning. QLoRA, GRPO, DPO, KTO, distillation. Reproducible runs on Unsloth Studio and torchtune.
- Agent architecture & red-teaming. Multi-agent topologies, RL-routed orchestration, threat models, cost-aware token budgets.
- Autonomous agent swarms. From a single tool-using agent to coordinated swarms of up to 300 sub-agents with deterministic guardrails, replayable runs, and human-in-the-loop checkpoints.
What I will not do
- Build something I cannot evaluate.
- Ship code I cannot sign.
- Take an engagement that requires lock-in for ongoing maintenance.
- Use customer data to train models outside the engagement scope.
- Promise outcomes I cannot measure against an explicit baseline.
Open-source & community
I publish boilerplates, MCP servers, and evaluation suites under permissive licences. I contribute to upstream projects when the engagement context allows. Code that improves the open-source baseline benefits every future client.
Booking
The fastest way to start a conversation is a 30-minute discovery call on cal.com/julien-compain. We talk scope, constraints, and acceptance criteria — you receive a one-page proposal within 48 hours.
Email : julien@novaquantix.tech· Generic inbox : contact@novaquantix.tech