Skip to main content

Cimiento — a governed AI-infrastructure control plane: intent XML → sense/plan/ground/compile/secure/govern-deploy/operate on Kubernetes, with machine-checked safety invariants (INV-22..32).

Project description


license: other license_name: cimiento-commercial license_link: LICENSE tags:

  • kubernetes
  • control-plane
  • ai-infrastructure
  • mlops
  • llm-serving
  • fine-tuning
  • agents
  • governance library_name: cimiento pipeline_tag: other

Cimiento — a governed control plane for AI infrastructure

Cimiento

A governed control plane for AI infrastructure. You declare what you want in a short intent file; Cimiento senses the live cluster, plans the endpoints, proves the plan fits the measured capacity (or abstains), compiles hardened Kubernetes manifests, applies them through a governed gate with a tamper-evident journal, and then operates them — watching SLO, drift, and cost. Every stage is guarded by a machine-checked invariant with teeth: disable the guard and its check goes red.

It is not a YAML templater and it is not an "IaC copilot." It is a small, auditable reasoning layer that refuses to ship a plan it cannot justify against the cluster in front of it.

pip install cimiento — pure-Python, no heavyweight runtime deps. 90-day free trial, then a signed key. See Licensing below.


Why it exists

Most infra automation fails open: it renders a manifest and hopes the cluster can run it. Cimiento fails closed. Grounding runs a quantum-inspired QUBO placement over the capacity it actually measured; if the workload does not fit, the deploy abstains rather than half-applying. The apply step is gated (dry-run → signed token → journal → rollback), so every change is attributable and reversible.

One governed pipeline, three verticals — each stage checked by a named invariant

Proven live, not on a slide

Cimiento measured results

Everything below was run end-to-end on a 4×A100-80GB RKE2 cluster and the machine-checked test suite — no mock-ups:

Result Detail
Qwen3-8B served live governed deploy … --live --apply → 4 pods, /v1/chat/completions → HTTP 200
LoRA fine-tune to completion Qwen3-1.7B, 30 steps in 63 s on 1×A100, train_loss 1.519 (governed K8s Job)
Governed agent CPU agent server calling the cluster's own model; invoked an allow-listed tool only
17.8× placement speedup GPU QUBO annealer vs CPU @ n=512 (grounding)
11 invariants with teeth INV-22..32 — each guard is falsifiable and the suite injects the bug
131 tests green full suite, including the licensing crack-defense tests

The pipeline

intent.xml → SENSE → PLAN → GROUND → COMPILE → SECURE → DEPLOY → OPERATE
             (INV-23)(INV-22)(INV-24)(INV-25/27)(INV-27/29)(INV-26/30)(INV-28)

One pipeline compiles three verticals from the same intent grammar:

  • Serving — vLLM / TGI / NIM, GPU-placed, readiness-gated, ingress + Service.
  • Fine-tune — a PEFT-LoRA run-to-completion Job (reuses the serving image, no extra pull).
  • Agents — a stdlib agent server with a hard tool allow-list, model id auto-discovered.

Quick start

pip install "cimiento[k8s]"

# see what would be deployed (no cluster changes)
cimiento plan     examples/policy-rag.xml

# prove it fits the live cluster, or abstain
cimiento ground   examples/policy-rag.xml

# governed apply: dry-run gate → signed token → tamper-evident journal
cimiento deploy   examples/policy-rag.xml --live --apply --yes

# run every machine-checked invariant, with teeth
cimiento verify --teeth

A minimal intent:

<intent name="policy-rag" useCase="serving">
  <model repo="Qwen/Qwen3-8B" task="chat"/>
  <slo latencyMsP95="1500" minReplicas="1"/>
  <gpu type="A100" count="1"/>
  <security airGap="true" dataResidency="eu"/>
</intent>

Architecture

Cimiento is a thin, auditable layer over a reasoning kernel (antahkarana):

  • antahkarana.spanda.padartha — the typed ontology categories the plan is expressed in.
  • antahkarana.spanda.qbackend — the QUBO annealer used for grounded placement (CPU or GPU).

The control-plane logic (sense/plan/ground/compile/secure/deploy/operate, the invariant checker, the CLI) lives in cimiento. Only the public verification key of the licensing system is shipped; keys are minted out-of-band by the vendor.

Licensing

Cimiento is commercial software with a built-in 90-day free trial. After the trial, a signed Ed25519 license key is required. The trial and key checks are pure-Python, zero-dependency, and resistant to clock-rollback and state-tampering (monotonic high-water mark + checksummed, redundant state).

cimiento-license status                 # trial days remaining / license state
cimiento-license install <KEY>          # install a signed key

To obtain a key, contact mohitdeepaksoni@gmail.com.

License

Proprietary / commercial. See LICENSE. Not open source.


Cimiento — Spanish for "foundation."

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

cimiento-0.1.0.tar.gz (73.2 kB view details)

Uploaded Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

cimiento-0.1.0-py3-none-any.whl (83.6 kB view details)

Uploaded Python 3

File details

Details for the file cimiento-0.1.0.tar.gz.

File metadata

  • Download URL: cimiento-0.1.0.tar.gz
  • Upload date:
  • Size: 73.2 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/6.2.0 CPython/3.9.6

File hashes

Hashes for cimiento-0.1.0.tar.gz
Algorithm Hash digest
SHA256 74b455bb6f4a8f650064bd0895fb224229d5f88eba363148c825be0ec75286f4
MD5 b3645019a34e0ab70b7f298d7aeabd13
BLAKE2b-256 108c06d05fed293d889435ea7d14be0c84a4006ca80c9ffb06d6854ea759f61e

See more details on using hashes here.

File details

Details for the file cimiento-0.1.0-py3-none-any.whl.

File metadata

  • Download URL: cimiento-0.1.0-py3-none-any.whl
  • Upload date:
  • Size: 83.6 kB
  • Tags: Python 3
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/6.2.0 CPython/3.9.6

File hashes

Hashes for cimiento-0.1.0-py3-none-any.whl
Algorithm Hash digest
SHA256 9b2be9d53bc6138ee56176ff8b94897f3413f5e89199056957dd574188187d5b
MD5 a5844d82676739a9af2276155f671773
BLAKE2b-256 06f815071bbef7ee535169ffc221fc7607348077d942d76c6d9d1316581fc69c

See more details on using hashes here.

Supported by

AWS Cloud computing and Security Sponsor Datadog Monitoring Depot Continuous Integration Fastly CDN Google Download Analytics Pingdom Monitoring Sentry Error logging StatusPage Status page