Now that I've spent money on X Premium, let me reintroduce myself ☺️ Hi, I'm Vidushi. I just finished my 3rd year of BTech and currently work at Noveum. There isn't some grand story behind how I got here. I took PCM because my parents suggested it. Then I picked CSE (AI/ML) during admissions because, honestly, "AI/ML" sounded cooler than regular CSE. That's about the amount of thought I put into most life decisions: do it, overthink later, continue anyway. For most of my first and second year, I wasn't particularly serious about building skills. One day a friend asked me, "Tu BTech PPT banane ke liye karrhi hai?" and that comment hit harder than it should have. So I started trying to improve. Still trying, honestly. The funny thing is that I never really liked studying. I definitely didn't like math, and probability and statistics made no sense to me. Then AI happened. For the first time, I wanted to learn something for no reason other than curiosity. I'd bring printed papers to class, read tech blogs on my phone during lectures, and spend time learning AI concepts when I probably should have been paying attention in class. Some people probably thought it was performative. Maybe. But the people who actually sat with me never did. They encouraged me to learn more, and I'm grateful for that. So that's me: no master plan, no "I've wanted to do this since I was 12" story. Just someone who accidentally found a field interesting enough to stay awake for. Outside of AI, I have a completely reasonable obsession with cute stationery, stickers, and tiny things that make me happy. I also have a little corner of the internet where I put whatever I feel like building, writing, or documenting: vinspace.live It's basically a cutesy portfolio with no clear purpose. Much like its owner.

updated the portfolio, have a look again (lemme know what you think) vinspace.me

The lesson: a test only measures the mistakes you thought to include. If you're shipping agents, test your monitor on failures your agent has actually made. (Fair warning: only 26 real mistakes, so that 42% could really be anywhere from 24% to 57%.)

It also had to see everything. On the planted mistakes: Conversation only: caught 5 of 44 Tool calls only: 24 of 44 Both together: 39 of 44 And a clear checklist in the prompt helped more than giving the model extra time to think.

The humbling part: on the real mistakes, a plain rule with no AI caught 12 of 26. The model caught 11. For exact mismatches like a wrong number, plain code kept up with the LLM.

The model wasn't the problem. My test cases were. I'd planted big, obvious failures. The real ones were quieter, like a booking handled correctly with the wrong refund amount. My monitor had been studying for the wrong exam.

The setup: a 4B Qwen on my laptop reads airline support chats from τ²-bench and checks what the agent said against what it actually did. It looks for four slips: - "done!" when it failed - changes before the customer said yes - a wrong date or amount - extras nobody asked for

palindrome✨

btw i have a "normal" portfolio also: vinspace.me

mood was shit but atleast i looked pretty (also gave my 3rd NPTEL! god save me)

idk why and how 60 people follow me on gh probably these 60 people also don't know why they follow me 🫪🦖 ... (maybe all these are my alt gh accounts 😏 who is it to say)

it's been around 3 months of yoga damnnnn... i've never been this consistent with exercise ever in my life. and... i can do a lot of poses now 🫪 even plank for 40ish sec suryanamskar, warrior poses, inclined push-ups, bear hold push-ups i wanna do a proper push-up 😭

why did i ever study these topics? even implemented them, EVEN MADE NOTES!! 😭😂 have i ever used it irl? nope 😶 ig a lot of time, i study stuff just for the fun of it and they serve no real purpose 🥲 sad

some people 🙄💅🏻

i have a dino on my laptop 🦖

this felt like that spiderman meme (my melody edition 😭)

idk why people hate CD or FLA