Abhijeetsingh Meena

Building AI systems for scientific discovery
Built DataVoyager (1K+ researchers) w/ @allenai
Co-author, DiscoveryBench (ICLR’25)

[Abhijeetsingh Meena @Abhijeet24]

Built systems used by 1K+ researchers to automate data-driven scientific discovery.

- Primary builder of DataVoyager (ICML'24) in collaboration with @allenai
- Co-author, DiscoveryBench (ICLR'25)
- Contributed to Microsoft AutoGen (coding agents)

I work on LLM reasoning agents and multi-agent systems for scientific discovery.

Interested in working on hard, high-impact problems. Always open to connecting with others building in this space.

[Abhijeetsingh Meena @Abhijeet24]

@arpit_bhayani very few people says truth out loud. dumb people keep tweeting "if you are not running X concurrent agents, you're ngmi"

[Abhijeetsingh Meena @Abhijeet24]

@Sanjay_Ghemawat @JeffDean @OriolVinyalsML @quocleix @DiscoLoopAI Congratulations!

[Abhijeetsingh Meena @Abhijeet24]

@PriyanshuP1405 Are these screenshots of managing your own funds?

[Abhijeetsingh Meena @Abhijeet24]

@elliotarledge What opportunities do we have for contract or freelance work that needs cuda skills? i.e., do companies hire for optimizing their model training pipelines?

[Abhijeetsingh Meena @Abhijeet24]

@Tim_Dettmers I feel the same way, single agent sessions with decent level of human supervision is better productivity wise; two or three can be adventurous but depletes our bandwidth to check for such drifts.

[Abhijeetsingh Meena @Abhijeet24]

@chaotictransfem You can use @modal, they have a generous free tier.

[Abhijeetsingh Meena @Abhijeet24]

@ThisIsBhandari I forget about the visa issue, but you should be good as they reached you out.

[Abhijeetsingh Meena @Abhijeet24]

@ThisIsBhandari These are standard numbers for HFT, but one thing to note is this isn’t just about DSA, that kid might be good at variety of things including systems, performance, stats & prob, calculus etc.

[Ilya Sutskever @ilyasut]

if you value intelligence above all other human qualities, you’re gonna have a bad time

[Abhijeetsingh Meena @Abhijeet24]

@ViditOstwal Same. It has become a disadvantage. Even we are ready work in PST.

[Abhijeetsingh Meena @Abhijeet24]

Good piece on explainability in agents.

The real blocker in practice is trust. And trust still breaks when the “why” explanations behind failures feel more like post-hoc narratives than actual decision traces.

[Priyanka Kukreja @pri_kukreja]

https://t.co/47li9G0iUC

[Abhijeetsingh Meena @Abhijeet24]

Interesting, I agree that explainability is becoming an important adoption barrier for agents.

I still mostly trust agents for only smaller features, where I can verify each step in a sync loop to ensure they're actually doing what I intended.

One challenge is that "why" explanations often feel like post-hoc rationalizations rather than actual reasoning behind the model's decision process.

It's a bit like trying to infer intent from code alone. You can make educated guesses, but it doesn't always closely match the reasoning behind a design choice.

So the harder question is how grounded these explanations are in the real decision signals, versus a well-formed post-hoc narrative.

[Abhijeetsingh Meena @Abhijeet24]

@shookscooks Which api?

[Abhijeetsingh Meena @Abhijeet24]

@manthanguptaa Agreed. In my experience planning and running one, two or atmost three agents working on separate parts of codebase with thoughtful follow-ups is more fruitful than running 20.

[Jon Xu @xuster]

That’s a wrap on YC Startup School India. Bengaluru, you really brought it. We can’t wait to see what you’re going to build.

[Abhijeetsingh Meena @Abhijeet24]

Huge thanks to @ycombinator for organizing YC Startup School Bangalore today and bringing together 2000 builders under one roof.

The energy was real. Met people working across AI x security, AI x systems, AI x quant - lots of interesting overlaps.

Good reminder that a lot of interesting work is happening at the intersection of fields.

Also got to meet folks from Twitter/X for the first time in person, which was nice.

What stood out was how open everyone was. People were sharing what they’re building, asking questions, and just genuinely curious.

Really enjoyed it. Hoping to attend more events like this.