§ 00 — Independent practice

Available for new engagements

Production-grade AIinfrastructure
delivered.

Scroll ↓

Custom MCP servers, fine-tuned open-weight models, and autonomous agent swarms. Engineered in TypeScript, Rust, and C — Python reserved for training pipelines.

§BView live demo↓

· LIVE · tail -f mcp.eventssigned events · last 10 s

ed25519 · sha-512 · merkle

timestampleveltoolpayloadsignaturems

23:59:59.422SIGNEDmerkle.root()merkle leaves[1..12]0x3dbd3dbd3dbd3dbd207
23:59:59.446OKstack.inventory()rate.limit.allow0x8e868e868e868e86228
23:59:57.430SIGNEDkimi.stream()redis.checkpoint0x9b1b9b1b9b1b9b1b135
23:59:57.454STREAMclaude.tool_call()context.compress 12.4k→2.1k0xece4ece4ece4ece4156
23:59:57.457OKhitl.approve()build manifest #48370xf979f979f979f979103
23:59:59.443STREAMhitl.approve()telemetry.snapshot0x4a424a424a424a42263
23:59:58.465STREAMeval.replay()tool.permission.scope0x57d757d757d757d7251

verify · /api/playground/key · rfc 8032stream rate · 7 events / 10s · burst 24

p95 latency
142ms
slo 150 · live · 12 min
throughput
4,087tok/s
kimi k2.6 · nvidia nim
signed builds
2,841
ed25519 · 90 days
models active
07open-weight
+ claude opus 4.8
live · now
uptime99.97%
30-day window · eu-west · us-east

§ 01Evidence — live

Signed builds, replayable runs.

Every deliverable carries an Ed25519 signature, a Merkle-chained build log, and an Inspect-AI evaluation suite. The numbers below are pulled from the active engagement window.

§ 01 · Signed builds (90d)ed25519

Each artifact verifiable with public key at /api/playground/key.

§ 02 · MCP p95 latencyp95 · 12 min

147ms

slo · 150 ms

§ live

MCP playground

kimi k2.6 · live

Type a request — the agent will pick one of the signed tools (novasign, merkle, stack) and return a verifiable result.

§ 03 · Eval pass rateinspect-ai

Inspect-AI suites, pinned datasets.

§ 04 · Agent topologymcp · fan-out

        Claude  ┬─→ tool:novasign  ───→ Ed25519
                │
                ├─→ tool:merkle     ───→ root.sha256
                │
                ├─→ tool:stack      ───→ inventory
                │
                └─→ guard.hitl  ←──── human

§ 05 · Founder note

“Ship the binary the client can run on day 365 — not the slide deck delivered at the end of week 2.”

— Julien Compain · Founder & Lead Engineer

§ 06 · Models we fine-tune7 active

Claude Opus 4.8tool · 1M ctx
Kimi K2.6moe · 1T params
DeepSeek V4reasoning
GLM 5.1open-weight
Qwen 3.7multilingual
Gemma 4edge
Llama 4scout/herd

§ 02What we build

Four practices. One engineering bar.

Every engagement delivers a production artifact your team can run, audit, and extend independently.

§ 01mcp

Custom MCP servers

Connectors engineered for your stack — Claude Opus 4.8, Cursor, Windsurf, Cline, and headless agents. Schema design, authentication, observability, and production hardening.

→TypeScript & Python SDKs
→Streaming I/O
→Signed releases

§ 02fine-tuning

Open-weight fine-tuning

QLoRA, GRPO reasoning training, and distillation on the latest open-weight models. Reproducible runs on Unsloth Studio and torchtune — 70% less VRAM, eval-driven gates.

→DeepSeek V4 · Kimi K2 · GLM 5.1
→Qwen 3.7 · Gemma 4 · Llama 4

§ 03architecture

Agent architecture & audit

Multi-agent topologies, RL-routed orchestration, retrieval design, cost-aware token budgets and security review before deployment.

→Threat model
→Token & cost budget

§ 04agents

Autonomous agent swarms

Production swarms with deterministic guardrails, replay logs, sub-agent fan-out, and human-in-the-loop checkpoints. From single tool to 300-agent orchestration.

→Replayable runs
→HITL approval gates

§ 03Live workflow

From scaffold to signed production in seconds.

Every NovaQuantiX deliverable ships with Ed25519-signed builds, Merkle-chained logs, and replayable Inspect-AI evaluation suites. These are the actual commands.

nova@quantix : ~/my-tooltty.0

$ npx create-mcp-server@nova my-tool
  → Detected stack : TypeScript · Anthropic SDK
  → Scaffolding stateful tool with streaming I/O
  → Signing build with Ed25519
  ✓ ready in 1.8s · ./build/server.mjs (signed)
$ claude mcp add ./build/server.mjs --name my-tool
  → Verifying signature · Merkle root 0x4a9c…be21
  ✓ tool registered · 14 actions exposed to Claude
$ nova eval --suite production --replay
  → Running evaluations across 4 model variants
  ✓ all checks passed · p95 142ms · cost $0.014/run

§ 04Built for distribution

Run anywhere, deploy everywhere.

MCP servers run in a low-latency mesh across cloud providers, dedicated infrastructure, and on-premise hardware. Health-checked, signed, and replicated by default.

01Cross-region low-latency mesh
02Signed deployments (Ed25519)
03Merkle-chained audit logs

        ·   ·       ·             ·     ·
   ·          ·     ·   ·    ·         ·
        +------------+        +---------+
   ·    |  EU-WEST   |←══════→|  US-EAST|     ·
        |  paris     |        |  iad-1  |
        +------+-----+        +----+----+
               │                    │
               ▼                    ▼
        +-----------+         +-----------+
        | Ed25519   |         |  Merkle   |
        |  signing  |         |   root    |
        +-----------+         +-----------+
                   ╲           ╱
                    ╲         ╱
                  +-------------+
                  |  CLIENT     |
                  |  own keys   |
                  +-------------+
   ·       ·            ·          ·    ·       ·

fig. 04 — signed mesh, client-owned keys

§ 05Engagement model

From scoping to running system in six weeks.

Fixed-price phases, verified CI/CD gates, immutable build logs. Clear visibility on what is delivered and what comes next.

phase 01
Discovery & architecture
We map your data flows, agents, and risks. You receive a written Architecture Decision Record before any code is written.
Week 1 · ADR + budget
phase 02
Reproducible engineering
Each commit triggers an immutable build log, signed artifact, and Inspect-AI evaluation suite. Rust or C for performance paths, TypeScript for MCP, Python for training.
Weeks 2-6 · CI-gated phases
phase 03
Delivery & ownership transfer
We deploy, document, train your team, and transfer full ownership. Source code, keys, repositories, and evaluation baselines are yours.
Week 6+ · Source & runbooks

§ 06Technology stack

The components we ship with.

Each component is selected for performance, memory safety, and long-term maintainability. Python is restricted to training scripts where the ecosystem requires it.

Verify each release signature and Merkle root before deployment.

#	component	role
01	Rust	Performance-critical paths
02	C	Low-level control
03	TypeScript	MCP servers · agent interfaces
04	Python	Training pipelines only
05	Claude Opus 4.8	Tool use · 1M-token reasoning
06	Open-weight models	DeepSeek V4 · Kimi K2 · GLM 5.1 · Qwen 3.7 · Gemma 4
07	Postgres · Redis · pgvector	Stateful agents · vector & KV
08	Ed25519 · Merkle	Signed releases · audit logs

§ 08Generate · live

Get your one-page proposal in 30 s.

Type a brief — or start with /mcp, /tune, /audit, /swarm, /eu-ai. Kimi K2.6 drafts a fixed-price, scoped, signed proposal using NovaQuantiX's canonical ranges. Saved locally for your next visit.

§ ai · brief → proposalkimi k2.6 · nvidia nim

Your brief — one or two sentences

0/600·/ slash for templates·⌘↵ to generate

§ 07Get in touch

Ready to deploy production AI?

Tell us about your project. We respond with a one-page proposal — scope, budget, risks — within 48 hours.

30 min · response within one business day

Production-grade AIinfrastructuredelivered.