Civic technologist building AI tools that help systems explain themselves. Clew Suite builder, NACM 2026 speaker. I write about hackathons, product judgment, agent readiness, and learning in public.
@earlgreyhot1701d Thank you! 🙌 That brake is my favourite part of the code too... a budget alert only tells you after the money is gone, so I wanted something that stops the call itself. Would love to hear what you use for cost control in your builds!
Civic technologist building AI tools that help systems explain themselves. Clew Suite builder, NACM 2026 speaker. I write about hackathons, product judgment, agent readiness, and learning in public.
Your post changed my numbers! My Kiroween build, Dead Keys, is in progress and has an optional Bedrock layer. My first draft capped it at 500 calls a day for no real reason. After reading your point about alerts arriving after the money’s gone, I’m setting the brake from Nova Micro’s price instead: a $5 monthly budget, a cap of 3,000 calls a day, and 30 an hour per player, all checked in code before the call. The Budgets alert stays as the backstop, not the brake. I’m designing my civic agent the same way: a spend ceiling checked before every run, and if it’s hit, the run doesn’t start. Thanks for the push. 🙌
Seems well organized yourney to AWS chaos - just my perspective looks chaos maybe this short guide allow me sometimes to free my mind to open this direction. Until that moment I just relax fare away of giants.
@pengeszikra Ha, AWS chaos is the perfect name for it 😄 That's exactly why I wrote it... to give the chaos a shape and a first step. And there's no rush at all, the giants will still be there when you're ready. Hope it makes the door feel a little smaller when you decide to open it! 🚪
Great read! Especially liked the focus on real-world challenges like cost control, idempotency, event ordering, and evaluating AI agents. The emphasis on building production-ready systems rather than just learning AWS services makes this roadmap genuinely valuable. Thanks for sharing!
@gambhirsharma Thank you so much! 🙏 Those are the topics I wish somebody had warned me about earlier... cost, duplicate events and ordering bite way before the fancy AI parts do. Glad it came across as practical and not just another services list!
The Step 0 framing is the part I will keep coming back to — putting money and blast radius before any model call, and treating a budget alert as an alert rather than a brake. The token ceiling inside the code is a nice concrete version of that. The other detail that stood out was reading the stop reason: a max_tokens truncation being treated as a complete answer is exactly the kind of quiet bug that looks like a bad model. And moving evaluation to Step 2 instead of Step 5 makes a lot of sense — five questions you already know the answer to is a cheap habit to start on day one. Thanks for walking the roadmap and saying plainly what you did and did not test.
@onelifevibe This made my day, thank you! 🔥 You picked the exact details I was most excited about... the stop reason one especially. A truncated answer that looks like a complete one is the kind of bug that makes you blame the model for days. And yes, five questions you can grade on day one is such a cheap habit with a huge payoff. Glad the plain "here's what I did and didn't test" landed too!
@waqarsethar That's honestly the nicest thing to read... thank you! 🥹 And old school programmers are the best at this, you already know why the fundamentals matter. That's why I love sharing my learning curves here. If any part feels confusing, tell me and I'll happily break it down!
We're a place where coders share, stay up-to-date and grow their careers.
I appreciate the brake inside the code! Adding another safety mechanism to help control costs is top of mind when I build.
@earlgreyhot1701d Thank you! 🙌 That brake is my favourite part of the code too... a budget alert only tells you after the money is gone, so I wanted something that stops the call itself. Would love to hear what you use for cost control in your builds!
Your post changed my numbers! My Kiroween build, Dead Keys, is in progress and has an optional Bedrock layer. My first draft capped it at 500 calls a day for no real reason. After reading your point about alerts arriving after the money’s gone, I’m setting the brake from Nova Micro’s price instead: a $5 monthly budget, a cap of 3,000 calls a day, and 30 an hour per player, all checked in code before the call. The Budgets alert stays as the backstop, not the brake. I’m designing my civic agent the same way: a spend ceiling checked before every run, and if it’s hit, the run doesn’t start. Thanks for the push. 🙌
Seems well organized yourney to AWS chaos - just my perspective looks chaos maybe this short guide allow me sometimes to free my mind to open this direction. Until that moment I just relax fare away of giants.
@pengeszikra Ha, AWS chaos is the perfect name for it 😄 That's exactly why I wrote it... to give the chaos a shape and a first step. And there's no rush at all, the giants will still be there when you're ready. Hope it makes the door feel a little smaller when you decide to open it! 🚪
Keep up great work!
Great read! Especially liked the focus on real-world challenges like cost control, idempotency, event ordering, and evaluating AI agents. The emphasis on building production-ready systems rather than just learning AWS services makes this roadmap genuinely valuable. Thanks for sharing!
@gambhirsharma Thank you so much! 🙏 Those are the topics I wish somebody had warned me about earlier... cost, duplicate events and ordering bite way before the fancy AI parts do. Glad it came across as practical and not just another services list!
The Step 0 framing is the part I will keep coming back to — putting money and blast radius before any model call, and treating a budget alert as an alert rather than a brake. The token ceiling inside the code is a nice concrete version of that. The other detail that stood out was reading the stop reason: a max_tokens truncation being treated as a complete answer is exactly the kind of quiet bug that looks like a bad model. And moving evaluation to Step 2 instead of Step 5 makes a lot of sense — five questions you already know the answer to is a cheap habit to start on day one. Thanks for walking the roadmap and saying plainly what you did and did not test.
@onelifevibe This made my day, thank you! 🔥 You picked the exact details I was most excited about... the stop reason one especially. A truncated answer that looks like a complete one is the kind of bug that makes you blame the model for days. And yes, five questions you can grade on day one is such a cheap habit with a huge payoff. Glad the plain "here's what I did and didn't test" landed too!
I'm old school programmer , however learning through the mind and eyes of @mursalfk makes my learning journey smoothie..
@waqarsethar That's honestly the nicest thing to read... thank you! 🥹 And old school programmers are the best at this, you already know why the fundamentals matter. That's why I love sharing my learning curves here. If any part feels confusing, tell me and I'll happily break it down!