AI Infrastructure Reliability & Economics Platform (airun)

ai-infrastructure • gpu-finops • nvidia-dcgm • h100-sxm5 • mfu-profiler • opentelemetry • circuit-breakers • disaster-recovery • gke-daemonset • eks-daemonset • aks-daemonset

"airun measures and optimizes the cost, latency, quality, and reliability of AI workloads."

airun is a local-first AI Infrastructure Reliability and Economics Platform for Python 3.11+. It bridges the gap between physical silicon, data center electrical power, multi-model execution, and enterprise business continuity.

---

Key Features

• Intelligence per Dollar (IPD) & Watt (IPW): Core economic efficiency metrics: $$\text{IPD} = \frac{\text{Validated Output}}{\text{Compute Cost} + \text{Energy Cost}}$$ $$\text{IPW} = \frac{\text{Validated Output}}{\text{Energy (kWh)}}$$
• Hardware Accelerator Power Modeling: Electrical TDP specifications for NVIDIA H100 (SXM/PCIe), A100, B200 Blackwell, L40S, Google TPU v5e, AMD MI300X, and Apple Silicon with data center PUE modeling.
• The Efficient Frontier of AI: Multi-dimensional Pareto optimal frontier plotting models across (Quality vs Cost vs Latency) for dynamic model selection without regression.
• Eval-Driven Routing: Policy engine enforcing quality SLA constraints (Tier 1 $\ge 0.95$, Tier 2 $\ge 0.88$, Tier 3 Economy) with background shadow testing on cheaper candidates.
• The AI Breaker Box: Multi-provider resilience with automated circuit breakers (CLOSED, OPEN, HALF-OPEN) that detect provider outages, latency spikes, and quality degradation.
• Semantic Equivalence Mapping: Automatic translation of prompts, generation parameters, and tool/function calling schemas across OpenAI, Anthropic, Google Gemini, and Local formats.
• AI Disaster Recovery (DR) Drills (airun dr drill): Automated fault injection simulations measuring capability parity, quality retention %, latency delta, and cost delta during outages.
• AI-Aware Causal Incident Graph: Diagnostic graph correlating hardware issues (GPU Xid 79, PCIe Gen1 throttling), network fabric stalls, and NCCL barrier timeouts to wasted dollars.
• Executive Command Center (airun ui): Interactive Web UI featuring real-time KPI tiles, Pareto frontier visualizer, circuit breaker status, and trace DAG inspector.
• Zero-Friction Tracing: Universal @trace decorator and with trace() context manager.
• Concurrent Critical Path: Accurately computes critical-path latency across parallel tools using interval DAG scheduling.
• Privacy by Default: Automatic API key and sensitive token redaction. Prompt and completion contents are never stored without explicit opt-in.
• Ultra-Low Overhead: Measured in-memory span overhead is $<20\mu\text{s}$ and total SQLite persistence overhead is $<1\text{ms}$ per workflow.

---

3-Minute Quickstart

1. Installation

# Python SDK & CLI (PyPI)
pip install airun-profiler

# Rust Real-Time Node Telemetry Collector (crates.io)
cargo install airun-collector

2. Environment Health Check

airun doctor

3. Run the Instant Demo (Zero Setup)

Run an offline multi-step agent simulation without external API keys:

airun demo

Output:

+---------- AI Workflow Runtime Summary -----------+
| Trace ID        5664cdc8296e41d2ab0b1feae7f0d481  |
| Workflow Name   multi_agent_coordination_pipeline |
| Final Outcome   [OK] SUCCESS                      |
| Total Cost      $0.0310                           |
| Cost / Success  $0.0310                           |
| Total Duration  411.7ms                           |
| Critical Path   411.7ms                           |
| Total Tokens    12,170 (in: 10,900, out: 1,270)   |
| Model Calls     4                                 |
| External Calls  2                                 |
| Retries         0                                 |
| Failed Steps    0                                 |
+---------------------------------------------------+

+-------------------- Findings & Optimization Insights ---------------------+
| * [CRITICAL] 56% of total cost comes from step 'agent_researcher' ($0.017)|
| * [WARNING] Token bloat: prompt context grew 2.7x from planner to synthesis|
| * [INFO] Concurrent execution: parallel tools saved ~411ms sequential delay|
+---------------------------------------------------------------------------+

                               Top Cost Drivers                                
+-----------------------------------------------------------------------------+
| #   | Span Name        | Kind | Model / Target    | Cost (USD) | % Total | Duration | Tokens |
|-----+------------------+------+-------------------+------------+---------+----------+--------|
| 1   | agent_researcher | llm  | claude-3-5-sonnet |    $0.0174 |   55.8% |  140.1ms |  3,020 |
| 2   | agent_planner    | llm  | gpt-4o            |    $0.0054 |   21.4% |  100.2ms |  1,590 |
| 3   | agent_critic     | llm  | gemini-1.5-pro    |    $0.0049 |   19.5% |   90.4ms |  3,310 |
| 4   | agent_synth      | llm  | gpt-4o-mini       |    $0.0008 |    3.3% |   80.2ms |  4,250 |
+-----------------------------------------------------------------------------+

Execution Hierarchy:
Execution Trace
`-- [OK] [workflow] multi_agent_coordination_pipeline (411.7ms)
    +-- [OK] [agent_step] agent_planner_phase (100.3ms)
    |   `-- [OK] [llm] agent_planner (100.2ms, model: gpt-4o, 1590 tok, $0.0054)
    +-- [OK] [agent_step] agent_researcher_phase (140.3ms)
    |   `-- [OK] [llm] agent_researcher (140.1ms, model: claude-3-5-sonnet, 3020 tok, $0.0174)
    +-- [OK] [agent_step] agent_critic_phase (90.6ms)
    |   `-- [OK] [llm] agent_critic (90.4ms, model: gemini-1.5-pro, 3310 tok, $0.0049)
    `-- [OK] [agent_step] agent_synthesizer_phase (80.4ms)
        `-- [OK] [llm] agent_synthesizer (80.2ms, model: gpt-4o-mini, 4250 tok, $0.0008)

---

Instrumenting Your Code

1. Synchronous Functions

from airun import trace, set_span_tokens, SpanKind

@trace(kind=SpanKind.LLM, model="gpt-4o", provider="openai")
def agent_step(prompt: str) -> str:
    response = client.chat.completions.create(
        model="gpt-4o",
        messages=[{"role": "user", "content": prompt}]
    )
    set_span_tokens(
        input_tokens=response.usage.prompt_tokens,
        output_tokens=response.usage.completion_tokens,
    )
    return response.choices[0].message.content

with trace("customer_support_pipeline", kind=SpanKind.WORKFLOW):
    result = agent_step("Summarize ticket #4821")

2. Asynchronous Coroutines & Parallel Tools

import asyncio
from airun import trace, SpanKind

async def run_agent():
    async with trace("travel_agent", kind=SpanKind.WORKFLOW):
        # Parallel tools executed concurrently
        async with trace("parallel_booking", kind=SpanKind.AGENT_STEP):
            flights, hotels = await asyncio.gather(
                fetch_flights(),
                fetch_hotels()
            )

3. Zero-Code-Change CLI Profiling & CI/CD Integration

Prefix any Python execution with airun run and capture trace IDs deterministically:

airun run --trace-id-file .airun/trace_id ./agents/my_agent.py
airun report $(cat .airun/trace_id)

---

Comparing Runs (airun compare)

Benchmark two architectures or model swaps side-by-side:

airun compare previous latest

                    Trace Comparison: 84081c95 vs 6f1fe728                    
+----------------------------------------------------------------------------+
| Metric         | Run A (Baseline) | Run B (Optimized)|       Delta (B - A) |
|----------------+------------------+------------------+---------------------|
| Total Duration |          425.0ms |          177.0ms |   -248.0ms (-58.4%) |
| Total Cost     |            $0.03 |          $0.0007 | $-0.027764 (-97.4%) |
| Tokens         |            8,100 |            3,290 |              -4,810 |
| Retries        |                2 |                0 |                  -2 |
| Failed Steps   |                0 |                0 |                   0 |
| Quality Score  |             0.94 |             0.95 |               +0.01 |
+----------------------------------------------------------------------------+

---

CLI Reference

Command	Description
airun waste [latest|<id>]	Detect the 4 physical compute waste bottlenecks (Dataloader, NCCL, PCIe, Eager Mode) & financial bleed
airun golden-signals [latest|<id>]	Display 4-Layer Golden Signals hierarchy (Economics, Efficiency, Reliability, Infrastructure)
airun profiler trace --pid <pid>	Open-source profiler hook with hardware counter sampling & Airun Cloud ROI assessment
airun metrics [latest|<id>]	Display Executive Economics: Intelligence per Dollar (IPD), Intelligence per Watt (IPW), Energy
airun frontier	View Pareto Efficient Frontier across Model Quality, Latency, and Cost
airun policy [list|evaluate]	Manage and evaluate automated closed-loop remediation policies
airun dr drill	Run automated Disaster Recovery drill with provider failover audit scorecard
airun breaker status	Inspect live AI Breaker Box provider circuit breaker health
airun doctor	Verify local workspace, database health, and micro-overhead
airun init	Initialize workspace configuration and trace storage
airun demo	Run built-in simulated agent workflow
airun run <script.py>	Execute Python script with active profiler and instant summary
airun report [latest|<id>]	View runtime summary, diagnostic findings, and execution tree
airun compare [previous|<id1>] [latest|<id2>]	Compare two runs and inspect economic/latency deltas
airun trace list	List stored execution traces
airun trace show [latest|<id>]	Inspect execution tree hierarchy
airun export [latest|<id>] --format [json|otel-json]	Export trace in raw JSON or OpenTelemetry format
airun ui / airun serve	Launch interactive Command Center Web UI and REST API

---

Production Architecture & Durable Technology Stack

airun transforms AI infrastructure from opaque hardware spend into an accountable, measurable, and autonomic engineering discipline. Rather than treating components as mere utilities, each language and layer has a durable architectural role:

┌──────────────────────────────────────────────────────────┐
│                       TypeScript                         │
│               Control Plane / API Gateway                │
│                 + Executive Web Console                  │
└────────────────────────────┬─────────────────────────────┘
                             │ PostgreSQL State, Decisions, Metadata, Policies
┌────────────────────────────▼─────────────────────────────┐
│                         Pub/Sub                          │
│                      Event Backbone                      │
└──────────────┬────────────────────────────┬──────────────┘
               │                            │
┌──────────────▼─────────────┐┌─────────────▼──────────────┐
│            Rust            ││           Python           │
│    Real-Time Data Plane    ││      Intelligence Plane    │
│ (Ring Buffer, DCGM, OTLP)  ││ (Correlation, Waste, MFU)  │
└──────────────┬─────────────┘└─────────────┬──────────────┘
               │                            │
               └──────────────┬─────────────┘
                              │
┌─────────────────────────────▼────────────────────────────┐
│                GKE / EKS / AKS Kubernetes                │
│             Production Execution Environment             │
└──────────────────────────────────────────────────────────┘

Durable Architectural Boundaries

Layer / Technology	Durable Role	Core Responsibilities
Rust (crates/airun-collector, crates.io)	Real-Time Data Plane	Node/cluster telemetry collector, high-frequency GPU telemetry (DCGM/NVML), in-memory ring buffer (10Hz / 100ms), OTLP span ingestion, critical-path DAG engine, tri-state circuit breaker runtime.
TypeScript (packages/control-plane)	Control Plane	REST/gRPC API gateway, authentication/RBAC, organizations, projects, cluster configurations, billing, policies, and executive dashboard.
Python (src/airun)	Intelligence Plane & SDK	Developer-facing @trace SDK, OTLP span exporter, Time-Window Correlation engine, Physics of AI Waste diagnostics, MFU calculator, and CLI.
PostgreSQL (deploy/postgres)	System of Record	Source of truth for state, decisions, runs, policies, recommendations, and aggregated 1-minute rollups (never raw unaggregated telemetry).
Pub/Sub	Event Fabric	High-throughput asynchronous event backbone for batched telemetry, run completions, and waste alerts.
Kubernetes (GKE / EKS / AKS) (deploy/kubernetes, deploy/helm)	Execution Environment	Managed GPU node pools with nvidia.com/gpu tolerations and DaemonSets mounting /var/run/nvidia-dcgm and /sys/fs/cgroup.

---

The Crucial Bridge: Logical vs. Physical Tracing

Observability tools (LangSmith, Langfuse) only observe the logical layer (tokens, API costs). Infrastructure monitors (Datadog) only observe the physical layer (average GPU utilization).

airun's category moat is the Time-Window Correlation Bridge connecting the logical trace span directly to physical silicon.

Run the hardware waste diagnostic on any trace:

airun waste --hardware

Output:

+--- ! HARDWARE WASTE DETECTED IN TRACE: 5ff3390b470d49dc93de5ebb04d292aa ----+
| Workload         research_agent_workflow (8x H100)                          |
| Bottleneck       Dataloader Starvation (CPU/IO Bound)                       |
| Symptom          GPU SM active cycles dropped to 38.2% (idle stalls) while  |
|                  PCIe TX was idle (<400 MB/s)                               |
| Financial Bleed  $1,599.36 / week ($9.52/hr | $6,854.40/mo)                 |
| Root Cause       Host CPU data loading workers starved accelerator between  |
|                  training mini-batches                                      |
| Actionable Fix   Increase DataLoader num_workers=8, set pin_memory=True,    |
|                  and pre-fetch tensors                                      |
| Expected Impact  throughput_gain: +31%, waste_reduction: -82%,              |
|                  weekly_cost_savings: $1279.49                              |
|                                                                             |
|             Culprit Execution Spans in Critical Path                        |
| +---------------------------------------------------------------+           |
| | Span Name               | Duration (ms) | Cost (USD) | Tokens |           |
| |-------------------------+---------------+------------+--------|           |
| | research_agent_workflow |       334.7ms |    $0.0000 |      0 |           |
| | agent_planning          |       160.9ms |    $0.0000 |      0 |           |
| | planner_llm_call        |       120.4ms |    $0.0073 |   1800 |           |
| +---------------------------------------------------------------+           |
+-----------------------------------------------------------------------------+

Or diagnose workload-level economics:

airun waste --workload customer-support-agent

Output:

+---- Airun Workload Economics & Optimization Report -----+
| Workload         customer-support-agent                 |
| Monthly Spend    $84,210.00                             |
| Potential Waste  $17,430.00 (20.7% recoverable)         |
| Top Issue        42% of cost from researcher agent      |
| Root Cause       Large prompt context + expensive model |
| Recommendation   Route 73% of requests to cheaper model |
|                                                         |
|       Expected Impact from Optimization                 |
| +------------------------------+                        |
| | Dimension | Projected Change |                        |
| |-----------+------------------|                        |
| | Cost      | -31%             |                        |
| | Latency   | -18%             |                        |
| | Quality   | -0.4%            |                        |
| +------------------------------+                        |
+---------------------------------------------------------+

---

The 3 Evolutionary Phases

• Phase 1 — Developer Platform: Local-first Python SDK & CLI (@trace, airun report, airun compare, airun waste, airun frontier) with microsecond overhead and zero network dependencies.
• Phase 2 — Cloud Control Plane & Silicon Bridge: Multi-tenant SaaS architecture connecting logical traces to physical silicon via Rust Data Plane DaemonSet (deploy/kubernetes/daemonset-agent.yaml, deploy/helm/airun-data-plane/), Pub/Sub, PostgreSQL (deploy/postgres/schema.sql), and TypeScript control plane (packages/control-plane).
• Phase 3 — Autonomous Multi-Cloud Platform (10.0 / 10.0):
  • In-Kernel eBPF Fabric Tracing: Kernel tracepoints (kfree_skb, net_dev_xmit) detecting RoCE/InfiniBand packet drops and PFC pause frame storms.
  • Multi-Cluster Cross-Cloud Federation (airun cluster [list|overview|recommend]): Global GPU capacity aggregation and intelligent MFU-per-dollar placement across GCP GKE, AWS EKS, Azure AKS, and on-premise DGX SuperPODs.
  • Live Silicon CI Hardware Testing: Dedicated physical GPU test suite (tests/hardware/test_silicon_hardware.py), Kubernetes GPU runner manifest (deploy/ci/gpu-runner.yaml), and GitHub Actions workflow (.github/workflows/gpu-hardware-ci.yml). $$\text{Observe} \longrightarrow \text{Understand} \longrightarrow \text{Measure Economics} \longrightarrow \text{Find Waste} \longrightarrow \text{Recommend Optimization} \longrightarrow \text{Remediate} \longrightarrow \text{Learn}$$

---

AI Workload Laboratory

Explore representative runnable agent archetypes in examples/lab/:

python examples/lab/run_all.py

• chat_workload.py — Simple Chatbot
• rag_workload.py — RAG Retrieval & Synthesis
• tool_agent_workload.py — Tool Agent with Parallel Concurrency & Retries
• coding_agent_workload.py — Code Evaluation & Bug Fix Loop
• multi_agent_workload.py — Multi-Agent Coordination Pipeline
• comparison_workload.py — The Same Workload, Two Ways (Cost & Latency Regression Lab)

---

Documentation & Deployment Guides

• Deployment Guides:
  • Google Kubernetes Engine (GCP GKE) Deployment Guide
  • Amazon Elastic Kubernetes Service (AWS EKS) Deployment Guide
  • Azure Kubernetes Service (Azure AKS) Deployment Guide
• Technical Architecture & Specifications:
  • System Architecture & Distributed Data Path
  • Master Product Specification (SPECIFICATION.md)
  • Developer & AI Agent Guide (AGENT.md)
  • 5-Minute Quickstart Guide
  • Pricing & Accelerator Energy Economics
  • Architecture Decision Records (ADRs)
  • Automated Release & Publishing Runbook
• External Validation Kit (AIRUN-100):
  • Milestone AIRUN-100 Charter
  • Campaign Outreach Playbook
  • Pipeline Teardown Template
  • Quickstart Checklist
  • Feedback Form
  • Friction Log
  • Pain Ranking Matrix

---

License

Apache 2.0 License. See LICENSE for details.