kernell-os-sdk 2.5.0

First public release of an agentic runtime for M2M coordination. Machines coordinate, verify, and settle value autonomously.

🚀 Kernell OS SDK

🧠 What is Kernell OS?

Kernell OS SDK is an installable agentic runtime that executes, routes, and optimizes AI workloads automatically across multiple models and cost tiers.

It is not just a library to call LLMs.

It is a system that:

• Decides how tasks should be executed
• Optimizes cost, latency, and quality in real time
• Learns from production via telemetry
• Improves itself through a continuous data flywheel

---

💡 In One Line

Kernell turns AI inference into an optimized, self-improving system.

---

🧱 System Architecture (Layered View)

┌──────────────────────────────────────┐
│           Application Layer          │
│   (Agents, copilots, workflows)      │
└──────────────────────────────────────┘
                  ↓
┌──────────────────────────────────────┐
│        Policy & Decision Layer       │
│   (PolicyLite, risk, cost, routing)  │
└──────────────────────────────────────┘
                  ↓
┌──────────────────────────────────────┐
│        Execution & Routing Layer     │
│   (Router, fallback, decomposition)  │
└──────────────────────────────────────┘
                  ↓
┌──────────────────────────────────────┐
│        Model & Cache Layer           │
│ (Local / Cheap / Premium + Cache)    │
└──────────────────────────────────────┘
                  ↓
┌──────────────────────────────────────┐
│        Telemetry & Learning Layer    │
│ (Telemetry, labeling, datasets, FT)  │
└──────────────────────────────────────┘

---

🔥 Core Capabilities

🧠 Intelligent Routing (Policy Engine)

Automatically selects the best execution strategy:

• local → fastest, cheapest
• cheap → low-cost cloud models
• premium → high-quality models
• hybrid → safe fallback path

Decisions are based on:

• confidence
• risk
• expected cost
• latency constraints

---

🤖 Execution Engine

• Task decomposition
• Multi-model orchestration
• Automatic fallback
• Parallel execution support

---

💰 Cost-Aware Optimization

• Expected vs real cost tracking
• Budget enforcement
• Savings measurement (savings_pct)

---

📊 Telemetry & Data Flywheel

Every execution generates structured telemetry:

• routing decisions
• cost and latency
• success/failure
• policy signals

Used to:

• debug production issues
• build training datasets
• improve policy models

---

🔁 Continuous Learning Pipeline

Built-in tools:

• dataset generation
• labeling from real outcomes
• SFT dataset creation
• LoRA fine-tuning pipeline

---

⚡ Semantic Cache (L1 + L2)

• In-memory cache (L1)
• Vector database (Qdrant) (L2)

Reduces:

• latency
• cost
• repeated computation

---

🌐 Classifier-Pro API

• FastAPI server
• External policy decisions
• Rate limiting

---

🧪 Production-Grade Validation

• Containerized install validation
• Smoke tests (real execution)
• Chaos testing (failure scenarios)
• CI release gates
• Benchmark system

---

⚡ Quickstart

1. Install

pip install kernell-os-sdk

---

2. Basic Usage

from kernell_os_sdk.router import IntelligentRouter

router = IntelligentRouter()
results = router.execute("Explain quantum computing simply")

for r in results:
    print(r.output)

---

💥 Real Example (Value Demonstration)

Task:

"Summarize a 10-page document and extract key insights"

Without Kernell:

• Uses premium model directly
• Cost: $0.25
• Latency: 3.2s

With Kernell:

• Classifies as medium complexity
• Uses cheap + partial routing
• Cost: $0.03
• Latency: 1.9s

Result:

• 💰 ~88% cost reduction
• ⚡ ~40% faster
• ✅ Same quality (verified)

---

🧠 How It Works (Internal Flow)

Input
  ↓
PolicyLite → decides route (local/cheap/premium/hybrid)
  ↓
Router → executes plan
  ↓
Fallbacks (if needed)
  ↓
Result aggregation
  ↓
Telemetry capture
  ↓
Dataset + training loop

---

🧪 Validation Modes

🟢 Normal Mode (Release Gate)

Validates:

• install
• import
• CLI
• router execution
• telemetry
• policy
• failure-mode

---

🟡 Chaos Mode (Resilience)

docker compose --profile chaos up

Validates:

• degraded execution
• service failures
• fallback behavior
• system resilience

---

📊 Benchmarking

Run benchmark:

python scripts/benchmark_runner.py

Generate report:

python scripts/benchmark_report.py

Metrics:

• savings_pct
• latency_delta
• quality_guardrail

---

🔁 Data Flywheel

Production → Telemetry → Labeling → Dataset → Training → Better Policy

---

🧩 Use Cases

• AI copilots
• autonomous agents
• cost-optimized inference systems
• multi-model orchestration

---

🚀 Roadmap

• Fine-tuned policy model (LoRA)
• Auto-install model on init
• Production deployment tooling
• Advanced chaos testing (latency, partial failures)

---

🧾 License

MIT

---

⚡ Final Note

Kernell is not just an SDK.

It is a system for managing intelligence as a resource.