Software Developer @CreoWisTech | Building @TheDivSoup | Technical Writer | Daily Post on Web dev, AI Tools | DM/Mail for paid Collab: [email protected]
Joined June 2020
- Tweets39.8K
- Following197
- Followers26.7K
- Likes46.6K
Pinned Tweet
I’ve got hundreds of ideas, notes, and half-written drafts sitting on my laptop.
The problem was never what to post.
It was actually turning those ideas into a consistent posting system.
So I built a small workflow:
→ My ideas stay in my own content folder
→ Claude Code helps organize and turn them into drafts
→ Buffer keeps everything in one place
→ I review, edit, and add my personal touch
→ Nothing gets published without me
The goal isn’t to let AI become the creator.
It’s to spend less time managing content and more time actually creating it.
I wrote about the setup and how you can build the same workflow in ~10 minutes.
Blog link in the comments 👇
Keep the PDFs that break your pipeline. They're a useful test set.
How I'd use them with OXR:
- run your worst pages first
- compare with the 101s vs 144.7s baseline (OXR vs PaddleOCR-VL-1.6, one RTX PRO 6000)
- swap the layout model if your regions differ; recognition and reading order stay together
Why they matter: VLM feedback + rollout disagreement → clustered causes → targeted data for the next RSI post-training round. OXR is the first doc parser to use it.
Scoreboard: olmOCR-Bench, OmniDocBench, DocParsingBench, all first. Open-source speed, first too.
github.com/SoMarkAI/OXR
Stop forgetting.
You stop doing mental math.
Stop emergency orders at 11 PM.
Your phone just got a better memory.
1. Water? Tell it 20L, 2.5L/day → reminds you in 8 days.
2. Groceries? Every 3 days without thinking.
3. Medicine? Weekly injections. Tracked automatically.
3. Revision? Spaced repetition: nudged on day 1, 3, 7, 21, 60; actually learn what you study.
4. Netflix renewal? Reminder 2 days before so you can cancel or keep.
and much more...
Open beta this week.
Want to try? say 'hi' 👋
#buildinpublic
I recently fixed a tricky race condition in the Stop button of a streaming AI chat.
The interesting part? AbortController wasn't enough.
On the first message, the frontend doesn't have a sessionId yet. The server sends it later through RUN_STARTED.
So if we abort the stream immediately: abortController.abort();
we can kill the connection before receiving the session ID required to call the backend Stop API.
The solution:
→ User clicks Stop
→ UI shows “Cancelled” immediately
→ Keep the stream open
→ Wait for sessionId
→ Call POST /sessions/{id}/stop
→ Abort the stream
→ Ignore any late events
I also checked the Network tab in ChatGPT and Claude to see how their cancellation flows appear at the browser level.
Claude showed a separate stop_response request, while ChatGPT's captured flow showed the streaming request being cancelled.
It was a great reminder that: Stopping a network request ≠ necessarily stopping the work happening on the server.
I wrote the full breakdown with the race condition, implementation, code, timeline, and the reasoning behind the fix.
🔗 Article link 👇🏻
Ajay Yadav retweeted
it’s time to share all of the learnings in the public, what I’ve learned while working for the last 2 years as a software engineer.
today i visited my private repo, where I dumped everything I’ve learned so far, including the features I built, the architecture that developed after a few iterations, bug fixes... each and every piece is there in my private repo.
now ive AI to get those dumped goldmines to the public so that I can revisit those info; also, it will be great content for all of you.
it will be in the form of LinkedIn/x/Medium/freeCodeCamp articles, and those things for the feed will be shared as a single post.
hoping to post it from tomorrow.
are you excited?
it’s time to share all of the learnings in the public, what I’ve learned while working for the last 2 years as a software engineer.
today i visited my private repo, where I dumped everything I’ve learned so far, including the features I built, the architecture that developed after a few iterations, bug fixes... each and every piece is there in my private repo.
now ive AI to get those dumped goldmines to the public so that I can revisit those info; also, it will be great content for all of you.
it will be in the form of LinkedIn/x/Medium/freeCodeCamp articles, and those things for the feed will be shared as a single post.
hoping to post it from tomorrow.
are you excited?
Unpopular opinion: most "AI wrapper" tools you're paying $20/mo for are a single Claude/GPT system prompt you could write in 10 minutes.
Before subscribing to the next AI [X] tool, ask:
- Is this orchestration/infra I can't easily build?
- Or is this just a prompt + a UI?
If it's the second, go write the prompt yourself.
Stop saying "I know JavaScript" until you can answer these.
7 JS interview questions that eliminate 80% of candidates in round 1 👇
7/ Promise.all vs Promise.allSettled.
all() fails the whole batch if even one promise rejects.
allSettled() always resolves and gives you the status of each one.
This comes up a lot in "how would you fetch multiple APIs safely" type questions.
That's all 7.
If you can explain each one out loud without looking it up, you're not a beginner anymore.
Follow @ATechAjay. I break down real interview questions and system design every week.
Repost the first tweet so someone prepping right now sees it too.
Stop asking "what's the best AI tool".
Ask "what's my bottleneck."
- Stuck reading a codebase → Repomix + Claude
- Stuck on a blank page → Granola for notes, Claude for structure
- Stuck shipping UI → v0
- Stuck researching → Perplexity, then verify the source yourself
Tool second. Bottleneck first.
Ajay Yadav retweeted
I’ve got hundreds of ideas, notes, and half-written drafts sitting on my laptop.
The problem was never what to post.
It was actually turning those ideas into a consistent posting system.
So I built a small workflow:
→ My ideas stay in my own content folder
→ Claude Code helps organize and turn them into drafts
→ Buffer keeps everything in one place
→ I review, edit, and add my personal touch
→ Nothing gets published without me
The goal isn’t to let AI become the creator.
It’s to spend less time managing content and more time actually creating it.
I wrote about the setup and how you can build the same workflow in ~10 minutes.
Blog link in the comments 👇