Pranav H

@arnesh_11 Beta

Opening up 10–15 spots for researchers to beta test something I’ve been building: PaperTrace. Lit reviews are brutal. You spend more time hunting papers and stitching sources together than actually thinking. Weeks of work that shouldn’t take weeks. PaperTrace handles the whole thing: finding, organizing, synthesizing across 250M+ papers, in one place. If you’re in research or academia: → DM or comment “Beta” and I’ll reach out. You’ll get lifetime Pro access. All I ask in return: regular, honest feedback while we shape this into something great :)

@i__m_sid Me too! they accepted mine in 17 minutes. No way they did a deep check on the application right ?

Applied last night thinking "no way this gets reviewed fast in India". Got an acceptance email within an hour. YC is clearly moving quick on this one. The April 18 Bengaluru event is going to be insane, 2000 hand-picked builders in one room.

Just applied to YC Startup School India... ...and got accepted in under 1 hour. ⚡ Meanwhile, thousands of Indian founders are still waiting for a response while YC hand-picks for their first-ever event in Bengaluru. Feels surreal. If you're building something real, apply now before it's too late. This could be the spark before the YC Summer '26 batch. Who's in? Drop your application story below 👇 #YC #StartupSchoolIndia #YCombinator #IndianStartups #Bengaluru

@HarshitNay80531 @ycombinator Me too! would love to chat :) See you there

got into @ycombinator startup school will be in bengaluru on april 18 lately been deep in ML infra + LLM systems (kubeflow, latency, scale, breaking things in prod). if you're building something ambitious or would like to have a chat, let's connect 👀

@soumilrathi context

Starting a memory researcher group chat. If you’re into memory systems, personalization, or context engineering in LLMs & AI systems. comment “context” to join.

@ellen_in_sf literature review

Starting an AI Researcher group chat. The space is growing fast! Comment “literature review” to join.

@anant_hq Me too! See you there :)

got into yc startup school india hmu if you're coming! lets meet up :)

@VinayakRaizada Me too! see you there :)

Excited to be joining YC Startup School in Bengaluru as well as this opportunity!!!

@kakashiSamaaa @ycombinator Lets go! See you there :)

@ycombinator Got selected for @ycombinator Startup School India Excited to learn, network, and build alongside some amazing founders. If you’re attending, hmu Let’s go 🔥 #StartupSchool #YC #Startups #Bangalore

Every student accepted into Startup School India now gets $25k+ in AI and cloud credits. Apply, get in, and start building: events.ycombinator.com/yc-sus-india

Got accepted to YC Startup School LFG

@shashwatvalid just got in :) see you there

see yall soon, im in

@Chiragjoshi_12 Me too! see you there :)

I got accepted into YC AI Startup School. Who else is coming?

Published a new preprint: PlanCompiler. Instead of letting LLMs generate arbitrary workflow logic, PlanCompiler compiles typed execution graphs with static validation and deterministic execution. Goal: more reliable, cheaper structured LLM systems. researchsquare.com/article/rs-918…

Built a compiler that restricts LLMs to a typed node registry instead of free form code generation. 278/300 first-pass successes vs 202 for GPT-4.1. Key finding: partial constraints invite evasion. The planner finds the open surface. Writeup + repo 👇

@the2ndfloorguy Built a constraint compiler that gets 100% first-pass on LLM pipelines where GPT 4o baselines at 34% (300 task benchmark, 6 sets). Writing it up as an arXiv preprint now. github.com/prnvh/llm-code…

something i've been wanting to say for a while. if you built something fun, weird, or interesting - tag me. dm me. share it with me. when i started out, i would ship things and nobody would notice. it's a lonely feeling. and honestly it stops a lot of good builders from continuing. i got lucky. slowly things started getting noticed. somehow ended up with a decent following. now, if i can give someone's work a little push, why not. that's the least i can do. share it with me. i'll reshare, shout it out, try to get more eyes on it. just trying to be the person i wish i had back then. you deserve to be seen 🍻

My benchmark: 300 tasks, 6 sets, 50 tasks each. Set D: 10-12 node pipelines. Compiler: 50/50 (100%) GPT-4o baseline: 17/50 (34%) The gap grows linearly with pipeline length.

The failure isn't hallucination in the usual sense. It's structural drift, the model loses coherence across long generations. Column names invented inconsistently. Deprecated API patterns. Logically correct but broken pipelines. It gets worse the longer the pipeline.

The second focuses on structural uncertainty in multi step workflows. Instead of letting the LLM freely compose steps, I built a system that constrains generation through a predefined node registry and compile a validated graph before execution.

The first focuses on behavioral uncertainty; when a model should or shouldn’t respond. Here's what SVMP did: -Confidence gating -Escalation routing -Multi-tenant isolation -Atomic session locking Basically: controlling when the model is allowed to respond.