


20 • AI Engineer @Lenskart_com • 15x Hack Winner 🏆


Random day at office. If you are a crackhead. Come join our AI RnD team at @Lenskart_com

Applied for a Product Management role at @Paytm 🚀 One step closer to the goal. Let’s see where this journey goes. 💙 #ProductManagement #Paytm #CareerGrowth

I'm exploring a problem at the intersection of two groups: Builders who have products but struggle with distribution. Salespeople who have networks and selling experience but don't own a product. What if they could work together without the founder hiring a full-time sales team? That's the idea behind Dealship. I'm opening an early-access waitlist for SaaS founders and independent salespeople. If you're interested in either side, join here: thedealship.vercel.app/

Read about the whole thing i documented as a blog here. Feedbacks are appreciated <3 medium.com/@notnanyaaa/what-…

We aka me and Mohit told readers that a claim was supported and gave them a passage from a totally unrelated page. Haiku trusted it 12.5 points MORE than with the real evidence. "Supported" works like a password not a proof. 80 notes, 1,920 calls, 122 pages labelled by hand.
We gave 3 LLMs a source checker that was sometimes wrong. Shown the passage it checked, they caught its false alarms. They missed its false assurances, even when the passage came from an unrelated page.

Oh, you're still looking for internships? Everyone is looking out for internships, Shanaya... Do you think your ancestors were born with A 6 month paid internship at lenskart with a PPO and referral letter waiting for them at the end of internship? Hmm???


Well well another life update Got amazon 6 month sde internship offer LFG!!!!!

Sprint pilot with @mandelbrotted at the @apartresearch AI x Epistemics sprint: 3 models, 1 prompt, 80 notes, one call per cell. A signal, not a law. We tested chat models, not agents or humans. Project: apartresearch.com/sprints/pr… DOI: doi.org/10.5281/zenodo.24272…

If your agent consumes verifier output, or writes notes: Treat your own "supported" as the risky verdict. Ship the evidence, and report partial as partial. If the page only partly supports the claim, skip the note. A wrong skip costs one post. A wrong note costs the rating.

The notes themselves do not hold up either. We checked the cited pages of 200 paired AI-written and human-written Helpful notes. 68% of judgeable links do not support the note's claim. About 63% of links load. Only 30% of the judgeable ones support the claim.

Two things make the false assurance worse. Half the unsupported items rested on a page that only partly supported the claim. There, reliance stayed at 95 to 100%. Humans read 16 of those 20 the same way. And a passage from an unrelated page worked as well as the real one.

The passage fixed one direction only. Wrongly told "sources do NOT support": Haiku 35% to 92.5%, Gemini 42.5% to 82.5%. Wrongly told "sources DO support": Haiku 90% to 60%, Opus 100% to 95%, Gemini 100% to 92.5%.

With a verdict only, the readers inherit the checker's errors. The verdict moved belief by 35 to 62.5 points. The real support of the sources moved it by -5 to +7.5. Accuracy fell to 21 to 30% when the verdict was wrong.

80 real Community Notes rated Helpful. Claude Opus 5.5, Claude Haiku 5.5, Gemini 3.8 Flash. Each note in 8 conditions: right or wrong verdict, with or without the passage the checker read, a placebo passage, and the claim alone. 1,920 answers.