23,Product engineer → https://xitter.cf/t.co/AGcFXwr9oR 🇮🇳 🇨🇭 | 3X startup | Hire me 👉🏻✉️
Joined December 2022
- Tweets6.3K
- Following531
- Followers2K
- Likes12.2K
Pinned Tweet
JUST 365 DAYS AGO:
I was :
👉🏻Underweight(44kgs)
👉🏻Under confident
👉🏻no social skills
👉🏻Addicted
👉🏻not future oriented
Now:
👉🏻Somewhat healthy(60kgs)
👉🏻Can talk to anyone
👉🏻No addiction
👉🏻Focusing on myself and my career
👉🏻Building meaningful connections
You just need a WHY!
ankan retweeted
To all the vibe coders,
If you all having a long running background process for a task and there are several tasks waiting in the queue, please add a checkpoint after each successful task :) big learning!
Also please add a heartbeat on the process that can literally save you thousands of dollar 💵
And if there is retries in AI processing add checkpoint and save and retry should happen from that saved point not whole process again..
By doing this i saved soo much money to another vibe coded platform lately!
Down the line in future, i want to own a video games company or work for a video game company,not because its cool or anything but amount of fun it can add to humans life apart from all this weirdest things happening in the world. I would be super fulfilled as a human. No B2B SAAS can ever do what a video game can do
To all the vibe coders,
If you all having a long running background process for a task and there are several tasks waiting in the queue, please add a checkpoint after each successful task :) big learning!
Also please add a heartbeat on the process that can literally save you thousands of dollar 💵
And if there is retries in AI processing add checkpoint and save and retry should happen from that saved point not whole process again..
By doing this i saved soo much money to another vibe coded platform lately!
I watched “The playlist” on Netflix last sunday and since then i got addicted to system architecture i literally now on my free time go and ready Engineering blogs.
My recent favourites are :
👉🏻Legora’s Engineering blog
👉🏻Harvey’s Engineering blog
👉🏻Posthog’s Engineering blog (mainly the “How we made feature flags faster and more reliable”)
Crazy time to be alive i mean just think and solve the problem for implementation and help you think with more clarity we have our claude and codex 🙈
While solving any tech problem with AI agents just follow a simple rule to avoid AI slops (I hate AI slops)
Start with :
Goal -> Gap -> Scope -> Constraints
Thank me later, please dont use any free skills available for this!
Reason is rvery product have a different outcome it should be related to your own idea but ofc you can create a skill or .md file for this specific task.
Dont create slops
I was investigating a vibe coded platform and asked the founder if he had taken care of few problems because he was about to spend $60,000 per month for the next 3 months and i was surprised to find problems which can absolutely destroy and even take the cost up for the founder to 300x.
There were problems like :
Workload distribution
Tenant isolation
Onboarding capacity
Context selection
Memory and token consumption problem
Queue growth
Repeated work
Ordering
Retrieval latency
Ofc DB load
Model limits
Query performance
Storage growth
Cost per useful result
Accuracy
And so many more!
I askef him he told i built using claude 5.5 then i understood the problem was not with claude it definitely did a great job but the problem was context provided to claude it was not aware of all those stuffs because he just built it for the sake of building it as a MVP!
Now the founder took a month more before releasing another version and then start marketing it.
Dm me if you do want a little less slop 🎮
Shit man woke up with cold and slight fever 🤒 i have a big meeting coming up this friday kinda nervous also gotta prepare!
Confused , what should be the extension i get?
👉🏻 .si
👉🏻 .com
👉🏻 .in
not .ai cause too expensive lol!
Posthog is actually for IndieHackers or product builder i literally thought posthog will blow like anything in this AI era but not many people integrating it… Not affiliated by them. (Hope so i get 🙈) but anyways you can dm me if you wanna set it up on your SAAS too..
Track the drop off to find exact issue and improve it you can easily increase 30-40% of the sale just by tweaking stuffs a bit better way 🫵🏻
Problem: Immigration teams need to try MatterZero, but access shouldn’t be open to everyone.
Decision: Keep pilots invite-only and make requesting one simple.
Build: Request → admin review → access grant + invite. Checks and throttling run server-side; requests stay separate from grants in Supabase, modeled with Drizzle.
Whats next?
Integrating Sentry + Posthog 👀🎮
Yesterday at Matterzero i created the built invite-only, passwordless signin with supabase magic links and authenticator MFA for admin 👀
Just wanted to make pilot access easy while keeing it controlled in V1 so no public sugnup, no admin bypass 📱📱
Also updated the landing page to be more cooler and some SEO improvements addressing problems of immigration companies 🎮
Now that coding language is slowly becoming less relevant it is absolutely possible to build stuffs with languages you always wanted…
Is there any language you want to build upon?
I can go first its : Swift 🥹📱
This is insane but learnings for me 💀 :
• Isolate personal agents from work Slack and shared identities anyy 💀
• Require explicit confirmation before any proactive post of sensitive data should never be posting anything without permission 💀
• Use narrow temporary read-only access and revoke it after each jobb!!
• Force an audience check and keep full action logs for quick deletion (if late)
• Treat agent names and cross links as risks,recheck permissions after every change
Wohooo!
Building MatterZero a voice first agentic workflow system for immigration teams.
The business problem is pretty simple:
a lot of immigration ops still involve repeated calls, WhatsApp follow-ups, missing documents, inconsistent dates, manual CRM updates, and staff spending hours turning messy client information into a case-ready file.
I am trying to automate that entire prefiling layer.
The flow looks more like:
realtime voice → STT → conversation orchestrator → structured case state → document processing → contradiction detection → agent/tool execution → human review → audit/evals.
The interesting engineering problems are not "how to make AI talk."
They are :
- keeping structured state as truth instead of chat history
- multilingual voice + barge-in
- document + voice provenance
- idempotent tool calls and retries
- stateful workflow orchestration
- MCP-style integrations
- human-in-the-loop controls
- latency tracing + regression evals
- catching inconsistencies across CVs, passports, employment letters, and what the client actually said
The end goal:
turn an unstructured client conversation + a pile of documents into a **complete, consistent, review-ready immigration file** before the filing team touches it.
Starting with immigration but building the runtime in a way that can eventually support other regulated B2B workflows too.
Stay tuned 🥱
