Sabitlenmiş Tweet
Rajiv S
1.3K posts

Rajiv S
@rjv_im
Coding for fun and life. Building @mfstacktech, Founder @betalectic since 2014
India Katılım Kasım 2021
78 Takip Edilen144 Takipçiler

Over weekend - I am trying to compare Claude vs Codex to decide what to use with same tasks. And some learnings (I am not affiliated with any of the tools, I am just a developer who codes 10 hours a day)
- Too early to become a fanboy for any tool (I am biased towards Claude, but I am learning my lesson)
- Claude is slower overall: Compacts too often, behaves same for any type of task, overkill mode for anything
- Claude seems better for review
- I love Codex 5.5 medium speed and experience. But I fell into this trap - it sometimes completely drops the ball on some tasks, Claude handles any type of task.
- I like Codex for the push back it gives, feels a valuable tool
I use Codex 10%, Claude 90% - I have to use Codex more.
A costly mistake to do right now: Becoming a fanboy for any LLM.
Just explore. I would like to try any other tool if you suggest. I tried Gemma and Qwen in local - I think comparing those against above is not fair. I don’t know for which kind of tasks I should use a Local LLM yet. Would love to learn more.
A nice surprise was I built an AI Agent with tools and used just Haiku - it handled tooling pretty well.
English

Introducing - AI Agents for Mutual Funds
(Video: Onboarding Agent helping distributor collect pan, name, dob and verify it against records)
Feels like Mutual Fund workflows are entering a completely different phase
Distributors reaching out to AI bots for onboarding, transactions, servicing, reports, and investor follow-ups.
Fintechs embedding AI journeys directly inside their existing apps.
AMCs offering much faster onboarding, recommendations, support, and servicing experiences.
Not by replacing existing apps.
But by adding AI agents alongside existing workflows.
At MFStack, while continuing to improve UI/UX across our apps, we’ve started building AI-X — AI Experiences for Mutual Funds built over @Cybrilla and @ONDC_Official
Agents which can work through chat interfaces, APIs, SDKs, MCP Servers, and A2A integrations ⚡️
Also means solving real infrastructure and engineering problems around privacy, PII leakage into LLMs, India data residency, workflow orchestration, and interoperability, hallucinations.
Interesting times ahead 🚀
English

@MiladmoHQ @shadcn I like the bottom fade. Is there a component which gives that out of box? Fader?
English

Experimenting with Gemma4 and Qwen3.6
- A well maintainced fullstack app where I do active development
- Tools: Ollama and OpenCode
-----
- Qwen3.6:35b couldn't understand the problem and come up with solution, irrespective of how I steer
- Gemma4:26b understood the problem and came up with solution one shot, but couldn't code it. Went into infinite loop each and every time.
Exciting times. Loved OpenCode UI and Ollama Ecosystem though - pull LLMs like NPM packages.
Will try with smaller use-cases.
English

As of today, how I use Claude vs Codex
- Claude: Think, Design a Feature, Understand context of existing feature spanning multiple files/modules, Judge and then code. You basically need someone to think along with you, hopefully better.
- Codex: You already know the file(s), the change, you want something quick. You basically want someone with fast hands, write code.
Doesn't mean Codex can't handle multiple files, I recently refactored some legacy code, 400+ files to use one button everywhere - It's just type of the task.
I haven't taken Codex for ride for design/architecture type of tasks. But there are definitely some tasks where Codex find a simpler, scalable fix. And Claude keeps thinking, inventing when it's not needed.
Also, with Claude recently since Opus 4.6, one sessions one plan, if you get into plan mode again in btw - you are in for a ride.
English

Cursor eats lot of memory, I switched to @zeddotdev (love it) but it doesn't have good plugin/extension support as it's now from VSCode world. I use SonarQube heavily, so missing out on that.
Which editor hits right balance even if it's paid?
English

Opus 4.6 was slow, sometimes hangs and is lost and I don’t see much difference in 4.7!
The model behaves almost same as 4.7, forgets that it implement the plan and suggests the same plan! You have basically allow it to implement again for it to figure out it know. A bug which Claude couldn’t fix since 4.6!
No one faces this issue?
English

And it does wildcard domain support. Do we OSS this? Or make a super cheap affordable solution?
Rajiv S@rjv_im
Custom Preview Deployment Setup for @mfstacktech . One DO Server and several scripts, Cloudflare, Certbot.
English

Custom Preview Deployment Setup for @mfstacktech . One DO Server and several scripts, Cloudflare, Certbot.

English


@heyblake mfstack.com Whitelabelled Mobile and Web apps for Mutual Fund Distributors, Fintechs, AMCs of India
English

From @mfstacktech Dashboard you can customize your Investor Home Page - and also get a "close to real" preview render right there while configuring. And it's just the beginning..! The screen builder supports different layouts for desktop and mobile. (Video of the screen builder coming soon)


English

Another ui.sh before and after… not trying to blow anyone away with over the top flashy stuff, focusing on the little details that add up to make things look polished instead.
More work to do but definitely better if you ask me.

English


Adding Dashboards to @mfstacktech
Reference: @pacekit_
Checkout: ui.pacekit.dev for some stat/chart components

English







