Technical Architect by title. Still curious why things work the way they do. AWS Community Builder & Leader @awsugblr
- Tweets 9,275
- Following 1,912
- Followers 1,383
- Likes 31,612
Watch this @AMD ad. I love it not because it only gives importance to humans, but because it explains that AI can make it perfect while humans make it beautiful.
3
46
Avinash Dalvi retweeted
Friday reminder of a thing that bit me and will bite you too:
Your CloudFront certificate has to live in us-east-1. Doesn't matter where the rest of your stack runs. ACM in your home region won't work, and the error message won't be kind about it.
It's in the pipeline walkthrough here 👉 lckhd.eu/06JYmg
#AWS #CICD #AWSCDK
1
1
154
Was testing OCR on a doc-to-vector pipeline today, screenshots and images going into chunks for semantic search. Code was agent-written, and the query side just was not returning results for text pulled from certain screenshots. First thought was a logic bug in the code. Then I stopped and went straight to Tesseract's own online OCR tool to check if the problem was the package itself or my code.
Tested a screenshot with highlighted text, the blue hover-state you get on a menu item, and it failed there too, same as in my pipeline. Took a new screenshot without the highlight, and it worked fine. That is when it clicked: the highlighted or hovered state of the text was distorting what raw OCR could actually read.
Fix was adding image preprocessing before OCR, so the screenshot gets cleaned up first and Tesseract can pick up highlighted text reliably.
Whenever I get stuck, find the exact spot or the actual culprit first, whether that turns out to be a code issue, a product outage, or an AWS infra issue, then work outward from there. Left alone, an agent will often just keep refactoring the surrounding logic instead of finding that spot, which wastes time and can end up hiding the real issue even further. In this case, the culprit was never the query logic at all, it was the highlighted text breaking OCR upstream.
29
When ECS stops a task, it does not delete the network interface right away. It kills your app container first and keeps the network alive until the end, so the app has time to commit transactions and push its final log lines before the network actually disappears.
1
63
So Amazon Cognito is now a core skill, aws-auth, inside the Agent Toolkit for AWS. Coding agents can set up user pools and app clients, and troubleshoot sign-in flows using best-practice workflows built in.
Wiring a Cognito user pool was never the slow part for me though. Real decisions sit deeper, token expiry, refresh token rotation, which OAuth scopes each app client actually gets, MFA policy per user type. An agent following a best-practice template still has to pick those values for your specific app.
Get one scope wrong on an app client and it rarely fails during testing. It fails after someone already got in through it. Faster setup is good, but someone who actually understands the auth flow still needs to review what the agent picked, not just confirm it ran without error.
Does anyone here trust an agent to pick MFA policy on its own yet, or is that still a human call every time?
aws.amazon.com/about-aws/wha…
2
58
Company is going through restructuring right now, the kind where nobody really knows what happens next. Some people are clearly still processing the shock of it. I only recently got into this ecosystem myself, so watching this unfold feels different than it would if I had years here already. All I can actually control right now is what is in front of me, so I am just going to keep doing that and deal with whatever comes next when it comes.
Have you been through this anytime in career ? What is your experience
2
2
134
We need to do this often.
For anyone deep into AI right now:
Step away from the screen and do some hands-on work.
Work in the garden, build or fix something, do practical chores - tasks where you still have to think, just about entirely different things.
Best way to clear your head when your thoughts are stuck in an AI loop. 😵💫
1
2
54
I love command px0 . what a magic!
px0.ai is actually bawaaal cheez :) You will literally never go back to using any other editor.
3
76
Most become a better engineer advice is just another roadmap to check off.
This is something different - read code you didn't write instead of your own, and you don't even need a merged PR for it to teach you something. Going through other people's open PRs and issues, at different experience levels, teaches almost as much as shipping the fix yourself.
If you want to become the engineer every team wants on their codebase.
Do this:
Read 10,000 lines of code you did not write, in a real open source project, and ship one fix the maintainers merge.
Follow these steps:
- Pick one open source project you already use in your own work.
- Clone it and get it running locally before you read a single feature.
- Read the README and the contributing guide before you touch the code.
- Open the issues tab and claim one bug labeled good first issue.
- Trace that bug through the code until you can explain it out loud.
- Write the smallest change that fixes it without breaking anything else.
- Add a test that fails before your fix and passes after it.
- Open a pull request and explain your reasoning like the reader has no context.
- When their feedback confuses you, ask AI to explain what the maintainer meant, not to reply for you.
Reading your own code teaches you nothing new. Reading 10,000 lines someone else wrote and shipping a fix they merge is what teaches you to work on any codebase you are handed.
Come back and tell me how many review rounds it took before they merged your fix. Be honest.
African engineers must be globally production-ready.
1
3
72
Picture is from Lalbaug garden.
If you are interested in learning how to tell your story about AWS, don’t miss your chance because it’s limited. We @awsugblr will be doing a different-style meetup this time with limited people in Lalbaug garden. Link in comments for registration.
2
2
143
Was reading about MS Dhoni recently. The part that make me think was how he built around each player's strength instead of trying to fix their weakness, because nobody on a team is the rockstar at everything.
Realised that's basically what I'd already been doing without a name for it: my Misal pav team. one person wrote the cleanest code, one shipped the fastest, one saw everything from the customer's side, one actually understood scale. never tried to make any of them well-rounded, just let each one go extreme at their one thing.
1
6
171
Working on converting Word docs into embeddings for vectorization. One of the features needed content plus screenshots and images from the doc, not just text.
Did not want to jump straight to an advanced model or paid capability for this, first time doing it.
Started basic. ChromaDB running locally with SQLite. Built the chunking, ran it, chunks processed fine.
Queried with some questions. Text answers came back close enough, nearest match made sense. But screenshots, not accurate at all. Model was not picking the right image context.
Asked my MLE if we had solved something like this before. They pointed me to Azure Document OCR.
Before jumping to that, I wanted to try a basic Python package first, just to see what quality it gives on its own. Not because I thought it would beat Azure, just to know where the baseline actually sits.
This is the part I keep learning again and again. Start basic first, know exactly where it breaks, then reach for the advanced tool with real reason behind it, not just because it is the recommended one.
Anyone else default to the basic version first before pulling in the heavier tool?
2
89