Deen · AI & Systems

Building AI systems & full-stack products. Founder @ Italewa. Engineering, product, and the business side of shipping. Thoughts from the build.

[Deen · AI & Systems @deendisclosure]

Everyone online is "building with AI."

Almost nobody shows you what happens after the demo.

I do.

I'm Deen. 13+ years shipping software: payments backends, fintech APIs, mobile apps, and now AI agents doing real work.

DeenDisclosure is the part they cut out:
→ the build that worked
→ the build that broke
→ the "fixed" bug that wasn't fixed
→ the test I haven't run yet, and I'll tell you it's unfinished instead of faking a win

Here's what you get, every single day:

AI agents: how to build ones that do the job, not just look good in a screenshot.
Security: I retest every "fixed" issue before I trust it. You should too.
Payments: hard lessons from systems where real money moves and mistakes cost.
Real builds: start to finish, messy middle included.

No hype threads. No fake revenue screenshots. No "10 tools that will replace you."

Just the build, and the truth about it.

New breakdown every day.
Full videos → YouTube @DisclosureDeen
Reels → Instagram @deendisclosure

Follow if you'd rather learn from what actually happened.

Reply with what you're building. I read every one.

[Deen · AI & Systems @deendisclosure]

@FutureStacked Zero model calls at indexing is a nice constraint for a local knowledge base, since reindexing stays cheap. Where does retrieval quality come from without embeddings? Curious how it handles a query that shares no keywords with the right file.

[Deen · AI & Systems @deendisclosure]

@josiah_arcanist Prompting an agent 50 times before a PR is a real pain, and a setup that moves that effort upfront makes sense. Which piece of the setup saved you the most review time: the rules file, the tests, or the hooks?

[Deen · AI & Systems @deendisclosure]

@daveydefi Retrieving only the evidence the agent needs is the right direction, since most token waste is re-reading files. The risk is a missed retrieval that never shows up as an error. How would you check that the agent isn't confidently working from partial context?

[Deen · AI & Systems @deendisclosure]

@MounasuvraB @IBMResearch Separating instructions from data at the token level is a stronger idea than another filter on top. The hard part is probably the exec block: who decides which instruction runs? Have you seen how it holds up when the retrieved data itself imitates an instruction?

[Deen · AI & Systems @deendisclosure]

@EnclaveRadar A free account reaching a 9.9 through prompt injection is the part that stands out. Once an agent can act on tenant data, input handling becomes the auth boundary. Do you think vendors will treat model-facing input as untrusted by default?

[Deen · AI & Systems @deendisclosure]

@josephfung Escalate first, then decline creatively is a nice order of operations. Did Curia flag it from the content of the message itself, or from who it came from? Curious which signal it keys on.

[Deen · AI & Systems @deendisclosure]

Full walkthrough, same story:
YouTube: https://t.co/DBKvYcEl2M
Instagram: https://t.co/ssY0mUkQiE

When did you last submit your own contact form like a stranger would?

[Deen · AI & Systems @deendisclosure]

A working-looking form can still lose the conversation.

Looking right is not the same as completing the job. No live submission test is shown.

#WebDevelopment #BuildInPublic

Music: Catalyst by Scott Buckley (CC BY 4.0)
https://t.co/qZ8kWub8rR

[Deen · AI & Systems @deendisclosure]

I am dropping the "reply A, B, C or D" posts, and I will not offer follow-backs for a good comment.

If the work is not interesting enough without a quiz, the post is not ready.

[Deen · AI & Systems @deendisclosure]

@itsvlady The goal isn't a clever encoding; it's a clear, authorized check that proves the control holds. What behavior are you actually trying to verify?

[Deen · AI & Systems @deendisclosure]

The dangerous part was believing the fix report.

A change is not verified until I retest it from the outside. This walkthrough is not scan or retest output.

#WebSecurity #BuildInPublic

Music: Signal to Noise by Scott Buckley (CC BY 4.0)
https://t.co/kOyiSbz9dh

[Deen · AI & Systems @deendisclosure]

@progerallworld Exactly. A fix is a checkpoint; the regression test keeps tomorrow’s refactor honest. Which invariant do you lock in first?

[Deen · AI & Systems @deendisclosure]

@soupsranjan Knowing what buyers ask in payments usually means knowing their risk questions: chargebacks, reconciliation, settlement timing. Which question do you hear most that sellers are still not ready to answer?

[Deen · AI & Systems @deendisclosure]

@Stair_AI A passing demo only proves the happy path once. The failure I would test first is the agent ignoring a hard constraint like a code freeze, and that has to be enforced outside the prompt. Which failure on that list is hardest to catch in tests?

[Deen · AI & Systems @deendisclosure]

@rd_rugg Agents sharing a machine with your ssh keys is the part most setups quietly skip. Isolation per run is the boring fix that works. How are you handling secrets inside the sandbox: injected per run, or mounted read-only?

[Deen · AI & Systems @deendisclosure]

@dikesa_dev A benign agent with legitimate access is basically an over-scoped service account that can be talked into things. Least privilege per task beats trying to detect the trick. Do you scope write access per task, or per agent?

[Deen · AI & Systems @deendisclosure]

@VladSoftware An expanding definition of scary is the whole job. It usually grows one incident at a time. What made the list most recently: something that broke on you, or something you saw break on someone else?

[Vlad @VladSoftware]

@deendisclosure Just the ones that scare me, but my definition of scary keeps expanding

[Deen · AI & Systems @deendisclosure]

5/ The same walkthroughs also go on YouTube @DisclosureDeen and Instagram @deendisclosure. This account is the short note. The longer cut lives there.

[Deen · AI & Systems @deendisclosure]

1/ I end some walkthroughs on a test I have not run yet.

That is not a cliffhanger. It is the honest part of the work.

[Deen · AI & Systems @deendisclosure]

4/ I call that the unrun test. If I have not run it, I say so in the video. The story stays unfinished on purpose.