IMPORTANT NOTICE ⚠️:

grok has made me spammer or something (jalta h sala grok meri growth dekhkarr) i dont know, so my replies might not be visible to you, maybe this post too would not be visible.

so please spread this word 🙏, otherwise my network is done.

don't repeat the mistakes i did.
learning agentic ai? start with a framework from start.
otherwise you'll end up later learning these because industries require it.

finally...a new error after one whole day :)

As Always 😎😎😎

these are current measurements of my ai agent.
tomorrow or after-morrow, hopefully with migrated project, i'd show you a comparison and that would be v1.

i chose Langchain + LangGraph + Mem0, because CrewAI was more about multi-agent systems, LlamaIndex is for RAG agents.

look at the "position" field
that's the bugger.
context : this element is actually placed in bottom-center part.
the reason why you should understand and evaluate everything.

caught this while evaluating my rag pipeline and looking why its precision and recall are 0

LITTLE PROGRESS TODAY 🚀
as i said yesterday :

corrected tool_retry_rate logic [which led me to update tool registry],
added argument_correctness using fuzzy logic and
testcase_percent_passed which weighs different parameters on the basis of their importance.

corrected ui_rr

today's conclusion... kal milte hhhhh
the eval thing is done, just some minor tweaks.

and nvidia nim takes a lot of time mannn, but that's okay atleast they're giving us 120B model

uploading unofficial and partially broken results of my ai agent. which would converge later!!

kya mei zindaaa huuuuu kya mei zindaaa huuuuuu....
chal rahiii hainn saanseee beh rahaa h khooon
(translation : am i alive?)

completed making handmade dataset...
i think i have left loose points, but i'd cover them and increase the size to around 70 after hosting the project.

now i'd setup evaluators and benchmark it.

whatd you think ...should i add more handmade test cases???

so for every webpage im indexing 2-3k documents with metadata like id, color, size, tag, text, etc

comment on this, do you think this is reasonable??
do suggest better approach if you're from frontend or aiml background.

thanks 🙏

guess what im going throughh

the faster i improve my projects > the faster i can start applying > the faster i can get out of my suffering

life is so shiiiiittttttttttttttttttt

in engineering, over the time i have come to a realization that you struggle fitting to a new thing, framework, device, but the faster you fail - the faster you learn it and get to the result.

getting to the result is what matters.

Gave an ai-based ai interview at micro1
the questions:
1) protect wrong args
2) (what metrics) evaluate ai agent latency
3) evaluate ai agents perf.
4) async def and def
5) evaluating different layers like serving, db, tools, etc
6) type handling
7) explain root cause to pm

QUICK UPDATE

creating EVALSET is the toughest job, lot of manual effort and brainstorming.

but i've completed the task 70% and now the Query UI RAG pipeline and Web Search Tools are remaining.

As soon as i complete it, i'd be creating a pipeline for EVALUATORS.

Harsh reality, but true.

India, support your own companies first. Build your ecosystem, create jobs, invest in skilled developers, support innovation, and spread awareness. The future will change when we start doing the real work.

Don't pick fights with other countries until we're genuinely prepared to compete. If food becomes expensive, jobs disappear, and ordinary people struggle to survive, what good will these fights do? We need to build first.

Being called “Vishwaguru” means nothing without real work. A nation's strength is proven through innovation, technology, opportunities, and the quality of life it provides its people.

Please stay away from my posts if you're here only to blindly support Modi, Rahul Gandhi, Elon Musk, or Trump. I'm not here to worship any individual. I'm here to talk about India's future.

Read my previous post. My intention was to support India, not attack it. I want our country to become so strong that others think twice before underestimating us or trolling us.

We're not facing the worst of the crisis yet, which is why some people don't understand the risks. Don't blindly support anyone. Use your own mind, understand the bigger picture, and think beyond political loyalty.

Let's talk about the real issues.

When I was in school, much of the syllabus was already outdated. Even in engineering, I saw institutions more concerned with getting projects submitted than genuinely supporting innovation. How many talented students are being held back by a system that doesn't encourage them to build real products?

Meanwhile, I personally know people in the USA who dedicate their lives to building technology, solving difficult problems, and developing innovations that could save lives. Some projects may never succeed, but they still invest their time, money, and energy into trying. That culture of building, experimenting, and taking risks is one reason strong technology ecosystems emerge.

Healthy criticism is necessary. It's how systems improve. Blind praise doesn't create progress.

And yes, Ambani deserves credit for making affordable internet accessible to millions of Indians. That has made a real difference. But supporting an Indian company doesn't mean we should stop asking questions or demanding more innovation.

My point is simple: build India's ecosystem until we can confidently say we're ready to compete globally in technology, manufacturing, AI, and innovation.

Don't blindly support Modi. Don't blindly support Rahul Gandhi. Don't blindly support Elon Musk or Trump.

Support good decisions. Criticise bad ones. Support India by demanding better results.

I'm open to healthy criticism, and if any part of my argument is wrong, I'm willing to listen. My intention is not to divide people. It's to encourage everyone to think independently and focus on building a stronger India.

National strength isn't built through blind loyalty, trolling, or bans. It's built through real work.

🇮🇳 Jai Hind. Jai Bharat. 🇮🇳

just some very rough drawings.
something i do whenever i reach college late!! (everyday)

its not good but im learning

just migrated to pure pip...and im happy that nothing broke, tools ☑️ monitoring ☑️ logging ☑️ server ☑️
my decision for conda was just that its too good at managing dependencies and pip breaks things everytime.

but "naa jaane wakt kiiii marzi h kyaaa..."

1 kachoriiii 2 samosaaaa in pip walo ka tooooo....kya bharosaaaaaa!!

please someone give me advice on working with pip 🙏🙏
im working on anaconda + pip currently
langfuse is not on anaconda, so im getting it through pip, and it just breaks whole environment

i tried improving upon race condition in ollama since earlier it was loading both all-minilm and granite4.1, and tried evaluating it, but the results shocked me, average time increased rather than decreasing.

that's why measuring things are worth it!