Arjun Dev Singla

Software Engineer @Oracle. Interests in Machine Learning and Product Building.

[Arjun Dev Singla @arjunnn_1612]

Damn, this is great!

[Lucas Jin @lucashjin]

we just built git for video editing.

[Arjun Dev Singla @arjunnn_1612]

@sumitsaurabh927 The only character I wait for in marvel

[Arjun Dev Singla @arjunnn_1612]

@lingodotdev @sumitsaurabh927 I just published PaperSwarm — Multilingual research assistant.

https://t.co/xGdToO03rp

[Arjun Dev Singla @arjunnn_1612]

Finally Submitted my project for the @lingodotdev Hackathon. 💯

PaperSwarm end to end. Let's wait for the results😁.

Special mention to @sumitsaurabh927 for awesome management of this hackathon.

[Arjun Dev Singla @arjunnn_1612]

Other things that shipped:
→ Graphs auto-save to library when switching papers → Gap tiles show which paper each gap came from

Huge thanks to @lingodotdev — saved us from some truly cursed translations of "attention head"

[Arjun Dev Singla @arjunnn_1612]

Day 5 & 6 of building PaperSwarm in public 🧵

(I think I slept 8 hours in 48 — two hackathons running simultaneously, one interesting and one for the boss)

Here's what shipped anyway 👇

[Arjun Dev Singla @arjunnn_1612]

Click "View PDF" → we parse the arXiv paper. Every arXiv paper is 2-column.

We built a column detector

Left column first, full-width headers at their exact position, right column after. Matches how humans actually read papers.

[Arjun Dev Singla @arjunnn_1612]

Hardest problem so far: gap deduplication.

Day 5 tomorrow: PDF export, citation lineage, full UI translation.

Following along? Drop a research topic 👇

Do checkout https://t.co/9lNFMfmEZQ

[Arjun Dev Singla @arjunnn_1612]

Day 4 of building PaperSwarm in public 🧵

A multi-agent research synthesis tool that reads papers, finds gaps, and translates everything to your language.

Today the dashboard came alive. Here's what shipped 👇

[Arjun Dev Singla @arjunnn_1612]

Natural language search now works end-to-end.

Type in any language → system detects it → LLM generates 5 targeted queries → Semantic Scholar → results in your language → click a paper → graph builds.

No arXiv ID needed. Powered by @lingodotdev

[Arjun Dev Singla @arjunnn_1612]

@mshenshz We solve this with an LLM dedup pass in the reconciler before the graph is built. But the real trick is doing it after collecting all parallel results, not during so you have the full picture before you collapse anything.

[Arjun Dev Singla @arjunnn_1612]

Day 3 of building for @lingodotdev 🔨

Architecture drop.

PaperSwarm is built as a swarm of specialized agents — each doing one job, independently.

No giant monolith.
No single LLM prompt.
Just small workers collaborating through queues.

Here’s the system design 👇

[Arjun Dev Singla @arjunnn_1612]

@mshenshz A naive system gives you 40 redundant gaps. A good system merges them into 5 meaningful ones and knows which papers are solving each, which are making progress, and which haven't touched it at all.

[Arjun Dev Singla @arjunnn_1612]

@mshenshz When you analyze 8 papers simultaneously, every agent independently finds gaps in the research. But "limited scalability of attention mechanisms", "quadratic complexity of self-attention", and "computational cost at inference" are all the same problem described differently.

[Arjun Dev Singla @arjunnn_1612]

@mshenshz The hardest part is gap deduplication.

[Arjun Dev Singla @arjunnn_1612]

Open for suggestions, and architecture discussion everyone. Do checkout https://t.co/9lNFMfmEZQ.

[Arjun Dev Singla @arjunnn_1612]

The goal isn't summarizing papers.

It's mapping the research landscape around a paper:

• Similar work
• Method differences
• Unsolved problems
• Future research directions
More updates tomorrow 👀
#buildinpublic #hackathon