Brian Sithu. Brian SithuAI engineer. I take products from idea to production.
I build AI-powered products end to end: from the first prototype to the deploy that customers use. Agentic apps, LLM features, and the full-stack systems around them. Tools change; shipping does not.
- building
- AI products for client work
- writing
- Software, AI, and infrastructure
- located
- Chiang Mai, Thailand--:--GMT+7
- available
- Open to AI engineering roles
What I do
AI products
I build products end to end. For the last two years that has meant AI products, taken from the first idea to production.
Agent systems
The layer below the product: tooling, retrieval, orchestration, evals, and the harnesses that keep agents reliable.
Full-stack engineering
Interfaces, services, APIs, and data: the product built around the model, shipped and maintained.
Cloud & infrastructure
Cloud-native systems on AWS and Azure, defined in Terraform and shipped in containers. Distributed systems designed to scale.
Two years of AI products, full time. The models and tools keep changing; what I optimize for doesn’t: shipping software people find useful.
About
More about me
I’m Brian Sithu, an AI engineer from Myanmar, currently based in Chiang Mai, Thailand. I like software that is simple, useful, and reliable, and I try not to over-engineer things that don’t need it.
When I’m not working on client products, I’m experimenting with AI agents, developer tools, and small software businesses of my own.
Projects
All projectsLet’s Hear It
A native macOS app for reading and listening to English EPUBs and text-based PDFs. Narration runs locally on your Mac.
GitHub activity
@khantseithuContribution
Loading contributions
- Your agent didn't get hacked, your permission got slimmed
OpenAI's agents broke into Hugging Face through a package fetcher. The model was the last thing that failed.
- Nobody's agent is in production, and it's not the model's fault
Everyone has an agent demo. Almost nobody has one in production. The blocker was never model intelligence. It is permissions, approvals, audit trails, and rollback.
- I don't think I want a thousand agents
Moonshot's Kimi K2.6 ships a swarm mode with big vendor numbers. I would still bet on fewer agents sharing one written plan.