Specialization

Secure and Auditable AI Infrastructure on Kubernetes

Andrey Lesnikov · Senior DevOps Platform Engineer

AI calls, tool calls and GPU capacity treated as infrastructure: policy before execution, budgets before spend, evidence after every decision.

Read the guides Agent Security Contract
01

Projects

Three systems on this corpus.

Policy before execution, budgets before spend, and GPU admission still in progress — each claim scoped to its repository evidence.

Released, v0.2.x

Agent Security Contract

Fail-closed pre-execution gate for agent and MCP tool calls: a YAML capability policy with a pinned catalog digest, ALLOW / BLOCK / REQUIRE_APPROVAL verdicts (exit 0/1/2), CLI, GitHub Action and HTTP sidecar, and a JSON evidence envelope (policy_sha256, trace_id, commit_sha, envelope_sha256). The envelope digest is an integrity digest, not a signature. Results in docs/EVIDENCE.md are on this corpus of 66 scenarios (33 attack, 33 benign: 0 critical false-ALLOW, 0 benign false-BLOCK).

Case study ↗GitHub ↗MCP security guide ↗
Implemented in justrunme/ai-runtime-platform, off by default (INFERENCE_BUDGET_ENABLED)

Inference Budget Guard

Atomically reserves estimated cost (tokens × price) per tenant or key window before /v1/chat/completions, returns 429 when over budget and 503 (fail closed) on store or policy errors, then settles to actual priced usage. Scope on this test suite / this setup: estimates can differ from actual usage, and settlement can lag. Maturity matrix: Supported; docs/EVIDENCE.md rows say tested locally and re-run in CI.

Case study ↗budget-guard.md ↗OTel LLM guide ↗
In progress, not yet published

GPU Admission on Kueue

GPU admission control with Kueue quotas, cohorts and preemption for inference and training. Work in progress; repository and evidence will be linked when published.

Kueue vs Volcano vs KAI ↗
02

Principles

How the work is shown.

Public demo

Each project ships a runnable local demo or quickstart in its repo.

Reproducible tests

Claims come from generated files (docs/EVIDENCE.md, results/*.json), not hand-typed.

Evidence JSON envelope

Every decision can be exported as a machine-checkable record (Agent Security Contract envelope schema; ASC-shaped decision envelopes in the budget guard).

Pull requests with CI

Changes land through pull requests with CI checks. Each evidence row says whether it was tested in CI or only locally.

03

Related cases

Nearby evidence.

Case 27 · Flagship

Agent Security Contract

  • CLI
  • Action
  • Policy

Agent and MCP stacks expand the attack surface beyond prompt text: tool misuse, poisoned tool descriptions, argument injection, and unapproved high-impact actions run without a pre-execution gate

0 false-ALLOW · 0 false-BLOCK on this corpus

Read case ↗ GitHub ↗
Case 19 · Flagship

AI Runtime Platform

  • Python
  • vLLM
  • KServe

AI workloads often start as API calls but become production systems that need identity, policy, scaling, telemetry, and clear ownership

Every model call: policy + trace + budget

Read case ↗ GitHub ↗
Case 21

Cost and Token Observability

  • Meter
  • Owner
  • Budget

AI and cloud costs can grow quietly when usage is disconnected from teams, services, and deployment changes

Token spend visible before finance week

Read case ↗ GitHub ↗
Case 02 · Flagship

Cloud-Native AI Gateway

  • Gateway
  • Providers
  • OpenTelemetry

AI usage needs routing, policy, budget awareness, and provider resilience

One boundary: rate, budget, fallback, OTel

Read case ↗ GitHub ↗
Case 23

Zero Trust Service Mesh

  • Identity
  • mTLS
  • AuthZ

Internal traffic is often trusted by default, making lateral movement and policy gaps hard to see

East-west mTLS by default

Read case ↗ GitHub ↗
Case 15 · Pattern

Policy as Code Guardrails

  • OPA
  • Feedback
  • Promote

Security and platform rules are often discovered only after deployment or during reviews

Standards enforced in CI, not slides

Read case ↗ Proved in BlastGuard ↗
04

Guides

Deep dives.

05

Topic hubs

Topic hubs.

Focused entry points for agent security, LLM inference, GPU scheduling, AI infrastructure, and GitOps.

Agent and MCP security

Securing AI agents and MCP servers on Kubernetes: OWASP Agentic Top 10 controls, fail-closed tool-call gates, identity, egress and audit trails.

Open hub ↗
Available for meaningful infrastructure conversations

Build systems that stay calm under pressure.

Available for conversations

Talk infrastructure

Book a working session

Meet ↗

Evidence first

For deep dives

Explore the systems

All cases ↗