Director, Software Engineering | Actionable content for leaders in tech. Software engineering, engineering leadership, online business.

15+ years building production systems. Former Director of Engineering at Oracle, leading public cloud services and working with engineers from Microsoft, AWS, Google, Netflix.

My mission: help businesses and engineers build AI systems that create real leverage.

We're entering a world where AI agents do the work and humans direct. The question is which side of the gap you're on.

Tool user: Prompts ChatGPT, starts from scratch every time.

System builder: Has agents that research, write, and code - creating output while they sleep.

Most AI content teaches you to copy paste. I teach the patterns underneath - so you can build any system, for any problem.

One person with the right systems can do the work of 10. I help people become that person.

What I do:
→ AI Engineer: Community for AI engineers. Courses, resources, downloadable AI projects
→ GradientWork: Custom AI development for businesses
→ Newsletter: Free value every week

📩 newsletter.owainlewis.com
👥 skool.com/aiengineer
💼 gradientwork.com

Owain Lewis

5 AI frameworks worth learning in 2026:

(And when to use none of them):

Here's the dirty secret nobody talks about:

AI frameworks are wrappers over provider SDKs with extra features bolted on.

Many professional developers I respect actively avoid them.

The complexity. The over-abstraction. The breaking changes. It adds up fast.

I call this framework tax: the ongoing cost of maintaining someone else's opinions about how AI apps should work.

Start with the SDK. Build until you hit a wall. Then reach for a framework.

When you do, here's what's worth learning:

1. Pydantic AI (Python)

If you've used FastAPI, this feels instantly familiar.

Types, validation, contracts, retries on malformed responses. Works across all major providers.

Best for: Python engineers who already think in terms of types.

2. LangGraph (Python)

Treats AI applications as graphs, not chains.

Branches, retries, parallel execution, human-in-the-loop become first-class citizens instead of afterthoughts.

Best for: Production systems with complex workflow logic.

3. Vercel AI SDK (TypeScript)

Provider-agnostic with a clean developer experience.

Swap models and build agents with minimal code.

Best for: TypeScript developers building AI apps.

4. Google ADK

The first framework that takes polyglot teams seriously.

Official support for Python, Java, TypeScript, and Go with the same mental model. Ships with observability out of the box.

Best for: Teams building on the Google ecosystem.

5. Spring AI (Java)

Exposes LLM services through familiar Spring patterns.

The value isn't innovation. It's integration.

Best for: Enterprise teams already familiar with Spring. 

The uncomfortable truth?

The best framework is often no framework at all.

The goal isn't to master frameworks. It's to build reliable systems.

Sometimes a framework helps. Sometimes it just gets in the way.

What's your current recommendation?

__

Enjoy this? ♻️ Repost it to your network and follow Owain Lewis for more.

Want to master AI engineering?

Join thousands of professionals from top tech companies (Google, Amazon, Oracle) and level up your AI skills for free: https://lnkd.in/e7Ymdh_j

Founder, GradientWork | AI Engineer | I build custom software to make your business more profitable

5 AI frameworks worth learning in 2026

The easiest way to build AI agents in 2026.

(My micro-agent architecture explained):

I built an AI agent that searches YouTube, analyses thousands of videos in seconds, downloads transcripts, writes scripts, and uploads content.

It completely changed my workflow and how I think about "personal AI assistants". I use it every day. 

The whole thing? A folder with some files and about 150 lines of Python code.

Here's the architecture:

1. The folder IS the agent

A Micro Agent is a folder where files define everything:

AGENT md = who it is and what it can do
tools/ = scripts it runs
context/ = knowledge it reads
workspace/ = where it saves files

Point any terminal coding agent at the folder. Claude Code, Codex, Amp, OpenCode. Works perfectly with all of them. 

2. Scripts are a superpower

Software engineers have been writing scripts and CLIs for decades. Nothing new.

But when you give them to an AI agent, magic happens.

Terminal-based agents are exceptional at invoking scripts. This eliminates 90% of the complexity you get with setups like MCP.

No protocols. No complexity. Just stdin/stdout.

3. Skip the infrastructure

The pain of building AI agents is the harness. The UI. The orchestration layer.

There are hundreds of coding agents now. They already have perfect UX out of the box.

Don't rebuild what exists. Focus your energy on what makes your agent unique: powerful tools and rich context.

4. Version control native

Your agent is just files. Share the folder with your team and they can use the agent.

Track changes. Roll back mistakes. Collaborate on improvements. No deployment pipeline needed.

The craft is knowing what you don't need:

Most agents don't need vector databases. Or distributed orchestration. Or complex memory systems.

They need clear instructions, relevant context, and powerful tools that do useful things.

Is this suitable for every case? No. 
Is it suitable for multi user production situations? No
Is it a great way to quickly build useful agents for your own use? Yes.

What's your current setup for running AI agents?

---

Enjoy this? ♻️ Repost it to your network and follow Owain Lewis for more.

Want to master AI engineering?

Join thousands of professionals from top tech companies (Google, Amazon, Oracle) and level up your AI skills for free: https://lnkd.in/e7Ymdh_j

The easiest way to build AI agents in 2026

How I use Claude Code. 

7 tactics to make you a power user:

Claude Code is an AI agent that lives in your terminal. People from all industries are using it for far more than just writing code. Imagine it as the smartest personal assistant you've ever had. 

But the tool alone isn't enough. Here are 9 tactics I use daily:

1. Design before you build (think first)

Separate planning from execution. Use plan mode (Shift+TAB) to think through what you're building before Claude writes any code. What's the approach? What are the edge cases? Think first, code second. This is how senior engineers work. 

2. Break big work into tasks (plan the work)

I take a large spec/plan and automatically break it into smaller tasks to feed to an agent. Each includes what to build, why it matters, and what success looks like. Claude executes one at a time with a clean context window.

3. Iterate, don't one-shot (refine as you go)

No engineer writes code perfectly first time. So why expect AI to? 

After any first draft, I ask: "Review this and suggest improvements" or "What edge cases are we missing?" "How would you improve this code?".

4. Save commands you repeat (systematize patterns)

Use slash commands: I have a .claude/commands/ folder. /design starts a planning session. /review checks code before shipping. /commit writes proper commit messages. Write once, run forever.

5. Set up your project brain (configure once)

CLAUDE {dot} md sits in your project root. Claude reads it at the start of every session. Put your tech stack, code style rules, and things Claude should always (or never) do. You stop repeating yourself. Claude just knows.

6. Automate your standards (automate quality)

I use hooks to run tasks on every edit (tests, linting etc). Quality happens automatically. Problems get caught immediately, not discovered later.

7. Document more things (the mindset shift)

LLMs have no memory. Store important context in static docs. I keep a docs/ folder with SOPs, examples of good output, and rules for how I want things done. 

Claude is a powerful builder with amnesia. My job isn't writing better prompts. It's building docs and systems that make every session start strong.

The future of work = humans using AI agents to become superhuman. 

Which of these are you using yourself?

---

Enjoy this? ♻️ Repost it to your network and follow Owain Lewis for more.

Want to master AI engineering?

Join thousands of professionals from top tech companies (Google, Amazon, Oracle) and level up your AI skills for free: https://lnkd.in/e_KPUM5S

7 Claude Code tactics

Most AI agents never make it to production.

Because no one knows how to deploy them.

Everyone is vibe-coding great looking demos on YouTube.

But how do you take a demo and make it work for real customers?

Most production AI systems fit three patterns:

1. API service

User calls your endpoint, gets a response.

→ Chat APIs for customer support
→ Document analysis endpoints
→ AI-powered search

Use when you need synchronous responses or human-in-the-loop.

Pro tip: Stream the response. Standard request/response feels too slow for LLMs.

2. Scheduled job

Runs on a schedule. No trigger needed.

→ Daily report generation
→ Hourly data sync
→ Weekly newsletter drafts

Use for batch processing at regular intervals. Set it and forget it.

3. Event-driven

External event triggers your automation.

→ New email arrives, process and respond
→ File uploaded, analyse and categorise
→ Webhook fires, take action

Use when reacting to the world in real-time.

The pattern you choose determines your infrastructure.

On Google Cloud, for example:

API service → Cloud Run Service (scales to zero, scales to thousands)
Scheduled job → Cloud Run Job + Scheduler (runs to completion, stops)
Event-driven → Pub/Sub + Cloud Run (decouples source from processing)

AWS and Azure have equivalents. The patterns stay the same.

That's the gap most people miss.

The engineers getting hired aren't the ones with impressive demos.

They're the ones who can answer: "How do we give this to real customers?"

Which pattern are you building with?

__

Enjoy this? ♻️ Repost it to your network and follow Owain Lewis for more.

Want to master AI engineering?

Join thousands of professionals from top tech companies (Google, Amazon, Oracle) and level up your AI skills for free: https://lnkd.in/e7Ymdh_j

3 production AI pattersn

If you want to build AI applications, you need to know how to deploy them.

Most tutorials skip this part entirely.

They end with "and now you have a working chatbot." Running on localhost. Useless for real customers.

Here's what they don't show you:

- TLS certificates
- Load balancers
- Database connections
- Environment variables and secrets
- CI/CD pipelines
- Networking and DNS
- Async work and message queues
- Container orchestration

This is real architecture. Even if a platform team handles it at your company, understanding how software gets to production matters.

I deploy AI applications three different ways depending on the situation:

1. PaaS (Render, Railway)

Push code. They handle infrastructure.

Best for: Side/solo projects, MVPs, "just get it live today."

Trade-off: Costs add up quickly once you scale.  

2. Cloud Platforms (AWS, GCP, Azure)

Managed services with real infrastructure control.

GCP Cloud Run is my default: containerised, scales to zero, pay for what you use. AWS has ECS and Lambda. Azure has Container Apps. All three work - pick one and learn it deeply.

Best for: Client work, production apps, anything enterprise.

Trade-off: Steep learning curve. Each platform has its own approach to IAM, networking, and secrets. Budget real time to learn.

3. VPS (DigitalOcean, Hetzner)

Rent a server. SSH in. Run everything yourself.

Best for: Cost optimisation, GPU workloads, learning how it actually works.

Trade-off: You're the ops team. Security, backups, uptime: all you.

There's no universal right answer.

Shipping a side project? Start with Render.

Building for clients who need scale? Learn a cloud platform.

Want full control and lowest cost? VPS.

Is this the complete picture? No. Every app has different needs.

What's your go-to deployment path?

---

Enjoy this? ♻️ Repost it to your network and follow Owain Lewis for more.

Want in-depth tutorials on this stuff?

Join thousands of professionals from top tech companies (Google, Amazon, Oracle) and level up your AI skills for free: https://lnkd.in/e_KPUM5S

How to deploy AI apps to production

Qoder just launched their Teams Plan which brings agentic coding to enterprise software teams:

One of the challenges with agentic coding is working on larger real-world codebases. 

Qoder is an agentic coding platform that can search and reason across 100,000+ files, giving you deeper codebase understanding - not just individual files.. 

I use it as both a dedicated IDE and a CLI.

Here are the features worth knowing about:

1. Quest mode

This is a killer feature. Describe what you want. Qoder generates a spec, breaks it into tasks, and executes asynchronously.

You delegate, walk away, review completed work later.

This matches how we work as engineers: Plan > [tasks] > execute.

2. Repo wiki

Auto-generates documentation and keeps it in sync as code changes. Architecture overviews, module summaries, dependency maps.

3. Smart model routing

Qoder routes to the right model based on task complexity which makes token usage way more efficient. 

4. Teams Plan

→ $30/seat/month (minimum 2 seats)
→ 2,000 credits per seat
→ Centralized billing and admin console
→ SSO for team-wide access
→ Shared credit packages coming soon

Qoder is moving agentic coding from “nice to have” to serious infrastructure for enterprise software teams.

Check it out here: https://aisecret.co/Owain

If you want to get reliable data from LLMs, learn this essential technique:

The problem is simple:

LLMs return text. Your application needs structured data.

You ask for JSON. Sometimes you get JSON. Sometimes you get JSON wrapped in markdown. Sometimes you get a paragraph explaining the JSON.

Your code expects one format. The model gives you three.

This is why structured outputs exist.

What are structured outputs?

A way to force an LLM to return data in an exact format you define.

Instead of asking the model to return any old answer, you give it a schema. The model's output must match that schema. No exceptions.

When to use them:

→ Extracting information from text (names, dates, entities)
→ Classifying content into categories
→ Routing decisions in AI workflows
→ Any time downstream code expects specific fields

If your AI application needs predictable data from an LLM, don't rely on instructions.

Use structured outputs.

__

Enjoy this? ♻️ Repost it to your network and follow Owain Lewis for more.

Want to master AI engineering?

Join thousands of professionals from top tech companies (Google, Amazon, Oracle) and level up your AI skills for free: https://lnkd.in/e7Ymdh_j

Humanitec

I'm a principal engineer with oversight over the full breadth and depth of Humanitec's microservice stack. I've owned and led major projects including: designing and leading the bespoke CI offering used by all customer deployments executed in Humanitec; Driving strong API conventions and code generation across 30+ microservices to reduce bugs and integration issues; Introducing and leveraging reliable-outbox patterns in RabbitMQ to dramatically improve asynchronous reliability and reduce production incidents; Introduced formal incident management processes and guided development of multi-region disaster recovery; Led design and delivery of early LLM and general AI integration prototypes.

Principal Software Engineer

Score

I'm a core maintainer on the CNCF Score project that aims to provide an abstract application schema for supporting platform engineering APIs. We've held a booth at KubeCon London 2025, and ran multiple conference talks and maintainer workshops there.

Open Source Maintainer

Oracle

I worked as technical lead across OCI API Gateway, OCI Functions, and OCI Container Instances teams as well as liaising closely with the other teams in the Container Native Services group. My role included preliminary design, architecture, communication and risk mitigation as new projects and requirements are identified. This spanned not only customer-facing features being delivered but also big rocks emerging from oncall operations or inter-service requests. As a distributed systems SME I also worked to review architecture and proposals from other teams and worked regularly on shared tooling and libraries to assist teams in solving common problems.

Consulting Member Of Technical Staff

Built OCI API Gateway from the ground up into an OCI native service running globally in 16+ regions. Provided technical and operational leadership and guidance to teams working on cloud native technologies such as containers, Kubernetes, and customer-connected dataplanes.

Currently working as an architect for both OCI API Gateway and OCI Functions products.

Worked at the Cape Town Oracle office on Oracle Compute Infrastructure Classic (formerly OPC / Nimbula). Mostly worked in Python and Golang, on compute services and CI/CD projects.

Senior Software Engineer

Amazon Web Services

My third internship was spent improving the project I had built during my
previous internship. This involved increasing unit and integration test coverage,
improving scalability, redesigning in order to leverage new API's, and writing
more supporting documentation. Languages used: Python, Ruby

Software Developer Internship

As part of my second internship with Amazon Web Services, I worked on an
internal database server monitoring tool. This internship was valuable as I could
spend more time working on a successful project rather than spending more
time learning the ropes. Languages used: Python

Amazon Web Services is well known for their scalable cloud infrastructure and
services. I worked on internal EC2 tools as part of the Compute team. This was
invaluable experience at one of Cape Town's top companies contributing to
systems used by many people around the world. Languages used: Ruby on
Rails, Java

j5 International Limited

St James Software provides a suite of operations management applications for
heavy industry. I worked here for four afternoons a week performing bug fixing,
testing, and development.

Contractor

Worked on Warden, and open source monitoring and metrics application
encapsulating Sentry and Graphite python modules. Warden was designed to
run in Windows or Unix and monitor processes running on the system and to
relay events and errors to a central Sentry instance via SMTP.

University of Cape Town

Tutored first year computer science students Python and Java. This involved
helping students with their assignments, marking tests, giving advice, and being
available after hours to answer questions they may have.

Tutor

Principal Software Engineer at Humanitec | ex-OCI

Personal blog: https://bensblog.meierhost.com

• Experience: Staff engineer with diverse experience architecting, building, maintaining, and operating scalable networked distributed systems in public and private cloud settings.
• Responsibilities: Rapid and iterative proof of concepts; driving architectural decisions and review; identifying and solving technical risks upfront; building each teams’ understanding of formal distributed systems patterns and concepts. Leading knowledge share and blame-less post-mortems to drive technical excellence. Very familiar and comfortable with technical interviewing practices. Experienced at writing technical one-pagers and engineering proposals.
• Current Technology Focus: Most comfortable in Go, Java, or Python but adaptable to any appropriate languages and frameworks whether multithreading/multiprocessing, asynchronous, or event-loop based; Excellent knowledge of TCP, TLS, HTTP, web-sockets, GRPC; Security and operations focused; Strong understanding of distributed systems fundamentals; Strong understanding of Linux container interfaces and implementations.

taplio

The Battle

Owain Lewis

Ben Meier

Followers

Posts

Engagements

Likes

Performance ScoreOwain Lewis's LinkedIn Performance

Ready to grow your LinkedIn presence?

taplio

The Battle

Owain Lewis

Ben Meier

Followers

Posts

Engagements

Likes

Performance Score.css-goub6s{font-size:var(--chakra-fontSizes-md);font-weight:500;color:var(--chakra-colors-gray-600);display:block;margin-top:var(--chakra-space-1);}Owain Lewis's LinkedIn Performance

Ready to grow your LinkedIn presence?

Performance ScoreOwain Lewis's LinkedIn Performance