@itsvlady Honestly, I can't tell much difference between Medium and High, so Medium is probably enough. Have you tried Sonnet 5.5? It's been great for me, and you might not need Opus on High for most tasks.
I keep seeing "AI will replace software engineers."
I've watched friends and big companies go through layoffs, and a lot of people say AI is getting too good.
But I've also seen many of them find new jobs fast. Me included. I resigned from CBTW and got an offer from NAB pretty easily, plus a few others.
AI isn't killing engineering. It's changing how we work. People who don't adapt get washed out. The ones who do get better.
If AI is that good, why are Anthropic and OpenAI still hiring engineers?
Hey, Duc here 👋
I've spent the last 9 years building banking and scheduling platforms. A lot of things worked, a lot didn't, and I've learned plenty along the way.
Besides building Excalibase, I want to start sharing more of what happens behind the scenes.
Things like:
How I turn vague ideas into actual system designs
Lessons from building and maintaining systems in production (including the painful ones 😅)
How I use AI to think through features before writing code
My real Claude Code workflow, prompts, and feedback loops
Where AI genuinely helps, and where it just makes things more complicated
No guru stuff. Just real experiences, things I'm learning, and how I approach building software.
If you're into backend, system design, or building with AI, I'd love to have you around.
What would you like me to write about first?
It need to pass cicd pipeline for unit test and integration test first. Then I have schedule nightly run for e2e pipeline of all services.
Sometimes I will trigger e2e pipelines there are some big changes in one of those services.
And a bit of manual and AI check on real product after all
Hey, I'm Duc 👋
I'm a software engineer with 9 years of experience building global banking and scheduling platforms.
Since May 2025, I've been building Excalibase — an open-source cloud platform that brings your database, APIs, authentication, and app hosting together in one place.
No need to stitch together five different services. Just build and ship.
I'm building it solo, with my buddy Claude Code doing most of the heavy lifting. 😆
I'll be sharing my journey here:
What I'm building and how it works
How I use AI coding agents, harnesses, feedback loops, and workflows to build software
Why I choose one technical approach over another
Bugs, failures, and lessons learned along the way
If you're into databases, Kubernetes, developer tools, or building with AI, let's connect!
I'd love to follow what you're building, exchange ideas, and learn from each other. 🚀
@Galo2027 Isn't that what software engineers already do? Take vague requirements, break them down, plan the short and long term, and help win the deal. AI just makes the typing part faster.
The skill that made me better at both engineering and product wasn't a framework.
It was learning to break down vague requirements myself.
When a PO says "clients want X," I ask:
Why? What's the real problem?
What does our system do today?
What's the constraint: cost, infra, or deadline?
What's the smallest version that actually solves it?
What does the long-term version look like, and how do we get there?
Do this for 9 years and you end up learning the domain, the customer, and the system at the same time.
Writing code is the easy part. Knowing what to write is the job.
What's the question you always ask before building anything?
This is big for solo builders. I keep plans, diagrams and tickets in 3 different apps right now. Having them all linked in one place, next to the code, means less switching and less losing track. Going to try running Excalibase's next sprint fully inside it.
Starting to roll out today: Claude Whiteboards, Claude Sheets and Claude Data.
Think a plan through on a board, build a tracker in a sheet, and tell a story with your data.
So today I want to share how I use AI, mostly Claude, when working on Excalibase.
Excalibase is a big repo, with a lot of services running and talking to each other, basically trying to work like a small cloud provider. It's too much for one person to keep in their head, and too much for one chat session too. So the main thing I've focused on is memory.
Memory
All decisions are written down in Markdown, and every change starts with a Jira ticket that moves across the board, including the reason behind it.
claude-mem keeps track of what previous sessions did, so a new session can pick up the context instead of starting from zero.
When I tell it "no, this is not how we do it" or "this is the trade-off", that gets saved too. So over time, it gets better at following the decisions we've already made instead of repeating the same mistakes.
Agents
The main session acts as the orchestrator. It controls one or two coding agents and a separate reviewer, depending on the workload.
The reviewer gets a fresh checkout with no context from the coding agent. It has to come back with either OK or FIX, including the file, line, and what needs to be fixed.
Testing
Testing includes unit tests, integration tests, mocked tests for each service, and real end-to-end tests with everything connected.
There's also a nightly agent that runs the whole pipeline. If everything passes, we're good to cut a minor release.
A few rules I push hard:
AI needs evidence for everything. No guessing or making up numbers. Show me the actual run or log.
If Claude gets stuck or the results start getting worse than expected, I try a fresh session. I copy the relevant messages and give it only the context it needs. The results are usually much better because the new session isn't overloaded with old context or overlooking things buried in a long conversation.
I'm still the one making the decisions. I still code every day myself. Claude is getting better every day, so now a lot of the work is a loop of multiple agents working together and reporting back to me. I mostly get reports about what I need to pay attention to, what went wrong, or where something needs my input. If an agent finds something that requires a decision, it brings it back to me.
That's mostly it. Happy to share more details if anyone is interested in a particular part.
https://t.co/2D8F4k6OYP
Building Excalibase alone since May 2025, I've learned something about myself.
I want it to be perfect. It never is. Some days I ship a feature and spend the evening listing what's wrong with it.
And then the overthinking starts. Is this architecture right? Should I go back and redo this? Plans change, not because of users, but because of the voice in my head.
Being solo is fast. I decide, I build, done. No meetings, no debates.
But sometimes the debate is exactly what's missing. A second person who says "this is fine" or "no, that's wrong, here's why" would save me days.
If you've built alone, what do you do when you can't tell if you're improving the thing or just overthinking it?
Vibecoders don’t need to care about the language. Drivers don’t need to know how the engine works.
Engineers do.
If you want to move from vibecoding to building real things, learn what you’re choosing and why.
How one request goes through the excalibase engine
One engine serves every project. a request comes in on /{project}/graphql or /{project}/api/v1
The project comes from the URL path only. unknown project, refused before anything else runs
Then the JWT. keys come from our auth service's JWKS, and if the token names a project it has to match the path
Permissions for that role are cached. when you change one in Studio, provisioning sends a message on NATS and the engine drops the cache right away, no waiting for a TTL
The query gets parsed (cached too) and compiled into ONE SQL statement. the JSON is built inside Postgres and the permission rules are compiled into the WHERE, so a row you can't see never leaves the database
Then it goes to that project's own Postgres, with credentials pulled from provisioning's vault
Realtime is a separate path. the watcher reads Postgres changes and pushes them to NATS, and the engine checks every change against your select rule before it goes out on the websocket, with only the columns you're allowed
https://t.co/2D8F4k6OYP
@clementrog partly. git worktree per agent + every scratch/analysis file gets a temp name or a “TEMP, don’t use” header, reviewer skips those. Memory search still open but rule is past decisions only