kcearns / kevin-cearns v2026.10
Model card
§1 Model description
Kevin Cearns is a senior platform and AI engineer in Toronto with more than twenty years of building software. He founded Entuit in 2003 and still runs every engagement himself.
These days the work sits underneath AI: self-hosted inference on Kubernetes, agent fleets in microVMs, edge apps on Cloudflare, and the cost and security controls that keep them boring in production. He builds on a modern TypeScript stack and reaches for AWS CDK when a project needs it.
§2 Intended use
Designing and building the infrastructure AI products run on: inference platforms, agent runtimes, edge AI apps, and the observability, security and cost controls around them. Works best on problems where something has to run reliably and cheaply in production, not only in a demo.
Out of scope
- Prompt-engineering workshops
- Slide-deck AI strategy with nothing deployed
§3 Training data
- 2003 – nowFounder & Principal Engineer · EntuitSoftware, platforms and AI for growing businesses.
- 2025 – nowWriting in public · kcearns.techBuild-along series on agent fleets, inference platforms and AI economics.
§4 Capabilities
Scored by evidence: the number of posts on this site that exercise each capability.
- Infrastructure as code (AWS CDK)18 posts
- Kubernetes & platform engineering17 posts
- TypeScript systems17 posts
- AI agents & orchestration13 posts
- Security & isolation11 posts
- Self-hosted & hybrid LLM inference10 posts
- Observability & FinOps7 posts
- Cloudflare edge0 posts
§5 Evaluation
| Benchmark | Result | Domain |
|---|---|---|
| A Firecracker fleet for AI agents | Sub-second warm resume from snapshot | Agent infrastructure |
| Hybrid LLM platform on EKS | 8 parts, all reproducible CDK | Inference platform |
| AI demos on Cloudflare's free tier | $0 inside the free allowances | Edge AI |
| An integration pipeline that copes | 1 program for code + infra | Resilience |
| Hardening self-hosted LLMs | 5 OWASP LLM Top 10 risks addressed | Security |
| The agent control plane | 60–80% cost cut from hybrid routing | Orchestration |
| CI/CD as code with Dagger | 3 environments, one pipeline | Delivery |
| FlagSignals | A/B flags, experiments, billing | Product |
§6 Limitations & biases
- Asked for a blog post, has been observed to return an eight-part series with CDK code for each part.
- Strong prior toward measuring the cost of every request before agreeing a feature is done.
- Will route your workload to a local model whenever it can do the job; may be biased against paying for tokens.
- Knowledge of any given framework is only as fresh as its last deploy.
§7 How to use
import { Kevin } from "kcearns";
const kevin = new Kevin({
engagement: "platform" | "ai-infrastructure" | "agents",
via: "https://www.entuit.com",
});
const plan = await kevin.scope(yourProblem);
// → architecture, a fixed price, and a working system§8 Citation
@misc{cearns2026,
author = {Cearns, Kevin},
title = {kcearns.tech: Notes from the Inference Layer},
year = {2026},
url = {https://kcearns.tech},
note = {30 posts, ~144k tokens}
}