DuguetLabs
Vol. 01 № 2026 · 04
Status — Operating

An applied A.I. laboratory. Built for the practical minds behind tomorrow.

DuguetLabs designs, trains, and ships AI systems that hold up under production load. Inference at frontier scale, spatial perception, generative imagery, and agentic systems — five lines of work, sequenced and shipped as they pass internal evaluation.

See the catalogue How the workshop runs
§ 01

The Catalogue

Five lines of work, sequenced. One available today; the rest released as they pass internal evaluation.

  1. Plate I Live

    Inference as a Service

    An OpenAI- and OpenRouter-compatible endpoint serving frontier open-source and proprietary models on dedicated, sovereign infrastructure. Drop-in for existing SDKs; no migration.

    • Models25 models · DeepSeek · Llama · Mistral · Kimi · Cohere · Phi · Grok · GPT-5
    • Latencyp50 < 1s · streaming SSE
    • CompatOpenAI v1 · OpenRouter v1
    • RegionEU · US — your choice
    • Ships asKataleptic
    Use it now Documentation
  2. Plate II In preparation

    Spatial 3-D Detection

    Real-time 3-D scene understanding for indoor and outdoor capture — object detection, pose, depth, and segmentation in a single pass.

    Join the waitlist
  3. Plate III In preparation

    Image Generation

    High-fidelity diffusion models with controllable composition, consistent characters, and edit-grade outputs. Trained on permissibly-licensed corpora.

    Join the waitlist
  4. Plate IV In preparation

    Agentic Systems

    Long-horizon planning, tool use, and multi-step orchestration with observability and human-in-the-loop checkpoints — built for systems that take consequential action.

    Join the waitlist
  5. Plate V In preparation

    Agentic Applications

    End-user tools where the agent disappears into the work — research, ops, and back-office assistants delivered as finished products.

    Join the waitlist
§ 04

The workshop

i.

From first principles.

We build the smallest piece that solves the actual problem. Fewer abstractions, more measurements, and a clear paper trail from input to output.

ii.

Sovereign by architecture.

Compute, weights, and data stay where you can audit them. Self-host the entire stack, or use ours under a contract that forbids egress.

iii.

Open where it matters.

We prefer open weights and open formats; we contribute the generalisable parts back. Proprietary work earns its keep on results, not on lock-in.

iv.

Production-shaped.

Every release ships with throughput numbers, latency p50/p99, failure modes, and a runbook. If it cannot be operated, it is not finished.