DEV Community: Nilesh Arnaiya

AI for writing research papers in 2026: Bibby AI vs Anara vs Underleaf vs Overleaf (honest comparison)

Nilesh Arnaiya — Tue, 01 Sep 2026 11:23:39 +0000

Every "best AI for research papers" list treats these tools like they compete head to head. They mostly do not. They solve different stages of the same painful workflow: read the literature, write the draft, get the citations right, make the figures, then ship a poster and slides.

Pick the wrong tool for your stage and you end up paying for three subscriptions that half overlap.

This is a breakdown of where each one actually earns its place in 2026. I build one of these tools (Bibby AI), so treat the Bibby section as biased and judge it on the facts. The competitor sections are written straight.

The 30-second version
Tool Best at AI writing Live citation search Figures / diagrams LaTeX-native Starts at
Anara Reading + literature review Limited (analysis, not authoring) Yes, traced to your uploads No No Free / $20 Pro
Underleaf Turning anything into LaTeX Rewrite / shorten / elaborate Inline search No Works in/around editors Free / ~$4.99 annual
Overleaf Collaborative LaTeX editing AI Assistant on paid tiers Limited No generation Yes (the classic) Free / ~$16-21 mo
Bibby AI End-to-end authoring Agentic editor Yes Yes (generated in-paper) Yes Free tier + paid

Now the detail.

Anara: best for reading, not writing

Anara (formerly Unriddle, YC-backed) is a research reading and analysis companion. You upload your PDF library and ask questions, and it answers with clickable citations that trace back to the exact passage in your documents. That traceability is the whole point. It uses a RAG approach so it does not fabricate paper titles the way general chatbots do.

It handles more than text too, so PDFs, video, audio, and images can live in one workspace, and it connects to Zotero, PubMed, arXiv, and JSTOR.

Where it stops: it is not a writing or typesetting tool. It will not build your LaTeX document, format for a journal, or generate figures. It is the tool you use to understand the field before you write, and to sanity check claims while you write elsewhere.

Pricing: free tier (small daily word and upload limits), Pro at $20/mo, Team at $30/seat/mo, plus a high-end Max tier if you want the Deep Search agent. Use it if literature review is your bottleneck.

Underleaf: best for converting things into LaTeX

Underleaf is an AI LaTeX assistant focused on conversion. Point it at a scanned paper, a whiteboard photo, a PDF, or a spreadsheet, and it hands back compilable LaTeX with the math structure intact. Equation conversion accuracy is its headline claim, and it is genuinely good at it.

The clever part is delivery: it runs as a Chrome extension that works inside your existing editor (including Overleaf) or as a standalone web app, so there is no copy-paste dance. It also does inline citation search and one-click rewrite, shorten, and elaborate.

Where it stops: it is an assist layer, not a full research platform. No figure generation, no poster or slide output, no end-to-end project management for a multi-chapter thesis. Think of it as the fastest way to get non-LaTeX material into LaTeX.

Pricing: free tier that is actually usable, paid from around $4.99/mo billed annually. Use it if your main pain is "I have this thing and I need it in LaTeX."

Overleaf: best for collaboration and habit

Overleaf is the incumbent. It is the collaborative LaTeX editor most labs already live in, and it now ships a LaTeX-fluent AI Assistant on its paid tiers. If your co-authors already use it and you just want AI autocomplete on top of a familiar editor, it is the low-friction choice.

Where it stops: the AI is an add-on to an editor designed in the pre-AI era, not a workflow built around AI. No citation generation engine, no figure generation, no poster or slide pipeline. You pay mostly for collaboration and familiarity.

Pricing: free tier, paid roughly $16-21/mo depending on plan, with an academic discount for institutional groups. Use it if collaboration and inertia matter more than AI depth.

The honorable mentions
SciSpace and Paperpal: strong for literature discovery and language editing respectively. Paperpal in particular is good at grammar, readability, and submission readiness.
Writefull: LaTeX-aware language feedback that plugs into Overleaf. Narrow but good at what it does.
Elicit and Consensus: literature search and synthesis. Same stage as Anara, different flavor. Not writing tools.

None of these author the paper for you. They each own one slice.

Bibby AI: built for the writing stage end to end

Here is the biased section, so weigh it accordingly.

Bibby is a LaTeX-native AI platform built around authoring rather than reading. The idea is to keep writing, citing, figures, review, and presentation in one place instead of stitching four tools together. It currently serves 10,000+ researchers across 50+ universities in 50+ countries.

What it does:

Agentic LaTeX editor, with a Rich Text mode if you want to write like a Word doc and let Bibby produce the LaTeX underneath
Live citation search inside the editor that inserts the real \cite{} and the BibTeX entry, so no hallucinated references

Academic Diagrams, agentic figure generation with 6,000+ scientific icons (Bioicons + SciDraw) and one-click insert into your paper. Closer to a BioRender rival than a writing feature
Posters and presentations generated from the paper you already wrote, so the deck is a transform of your work, not a from-scratch project the night before the deadline

AI paper reviewer, a LaTeX to Word converter, and a live conference deadline tracker for when you need the boring stuff handled

Where it stops: it is not a literature reading tool the way Anara is. If your bottleneck is understanding 80 papers, pair Bibby with Anara or Elicit. Bibby owns the part where you turn understanding into a submitted paper.

Pricing: there is a free tier, so you can test the editor, citation search, and figure generation before deciding. Current pricing is on the site.

So which one should you actually use
Drowning in papers to read → Anara (or Elicit / Consensus)
Have material that needs to become LaTeX → Underleaf
Co-authors already on Overleaf, want light AI → Overleaf
Writing the paper, figures, poster, and slides start to finish → Bibby AI
Language and submission polish → Paperpal or Writefull

Most serious researchers end up using two: one for reading, one for writing. The mistake is paying for two that both only read, or two that both only write.

Try Bibby at trybibby.com if the writing side is your bottleneck.

Question for the researchers here: how many separate tools are in your paper-writing stack right now, and which one would you drop first? Curious where the overlap actually hurts.

I Was Spending 30% of My Research Time on LaTeX. Then I Found Bibby AI — and Got That Time Back.

Nilesh Arnaiya — Mon, 02 Mar 2026 08:10:56 +0000

I Was Spending 30% of My Research Time on LaTeX. Then I Found Bibby — and Got That Time Back.

A researcher's honest take on the AI LaTeX editor that quietly changed how I write papers.

Let me paint you a picture.

It's 11:47 PM. My conference deadline is in 13 hours. My methodology section is solid, my results are clean — and I am deep in a rabbit hole trying to figure out why my \bibliography command is throwing a cryptic I found no \citation commands error while my .bib file is sitting right there, staring at me, completely fine.

I lost two hours that night. Not to research. Not to writing. To LaTeX plumbing.

That was the moment I started seriously looking for something better. That's how I found Bibby — and honestly? I wish I'd found it years earlier.

Why LaTeX Is Both Necessary and Maddening

If you're a researcher, you already know LaTeX isn't optional. Journals expect it. Conference templates demand it. Your collaborators use it. The output is beautiful, the math renders perfectly, and nothing else comes close for typesetting a 40-page paper with 200 citations.

But the experience of writing LaTeX? It's like driving a Formula 1 car where you also have to rebuild the engine every 20 minutes.

The Bibby team clearly gets this. Their pitch — "Write research, not code" — hit me somewhere deep in my exhausted researcher soul.

What Even Is Bibby?

Bibby is an AI-powered LaTeX editor built specifically for academic researchers. Not a general-purpose writing tool with LaTeX bolted on. Not a glorified text editor with syntax highlighting. It's a purpose-built research environment where every single feature — from citation management to peer review simulation — exists to answer one question: how do we give researchers their time back?

It's used by 9,920+ researchers from institutions like MIT, Yale, Berkeley, Harvard, and Max Planck. That's not a marketing number thrown on a landing page — that's a signal that people with serious research workflows have found this useful enough to keep using.

Here's a deep dive into every feature that's actually changed how I work.

🤖 Feature 1: Bibby Chat — The AI Writing Assistant That Lives In Your Editor

This is the centerpiece, and it earns that position.

Bibby Chat is a context-aware AI assistant (Uses Gemini 3.1 Pro) and latest as models evolve and that understands your entire document — not just the line you're on. You can ask it to:

Draft a skeleton introduction based on your topic
Rewrite a paragraph to sound more formal
Generate TikZ figures from a description
Explain a cryptic LaTeX error in plain English and fix it
Suggest improvements to your argument structure

The "context-aware" part is what separates it from just pasting your document into ChatGPT. Bibby knows your LaTeX environment, your bibliography, your package list. When it suggests a fix, it's not guessing — it knows what's already in your preamble.

I used it last week to completely restructure my related work section. What would've taken me 3 hours of shuffling paragraphs took 25 minutes of back-and-forth with Bibby Chat.

📚 Feature 2: Smart Citation Search & Insert

This one might be my single biggest time-saver.

Before Bibby, my citation workflow was: open Google Scholar → find paper → copy title → search Semantic Scholar → grab BibTeX → paste into .bib file → manually write \cite{key} in my document → compile → pray the key matched. Rinse and repeat for every single reference.

With Bibby, I type what I'm looking for, and it searches across Semantic Scholar, CrossRef, and more simultaneously. It auto-generates the BibTeX entry with correct formatting and inserts the \cite{} command directly into my document with one click.

It also does something I didn't expect: it suggests relevant papers based on what I'm writing. Like, I'll be drafting a paragraph about transformer attention mechanisms and Bibby will surface three papers I hadn't considered. That's not just saving time — that's actively improving the quality of my literature coverage.

200 million+ papers searchable. One click to cite. No more tab-switching.

🧮 Feature 3: AI Table & Equation Generator — Snap → Perfect LaTeX

I want to tell you about the moment I fell in love with Bibby.

I had a handwritten equation from a whiteboard session with my advisor. A gnarly multi-line thing with summations, subscripts, and a matrix embedded in it. I took a photo. I uploaded it. Bibby gave me publication-ready LaTeX in about 4 seconds.

Four seconds.

The Equation from Image feature converts handwritten or printed math to LaTeX through image upload. It handles:

Multi-line equations with align environments
Matrices and vectors
Complex summations and integrals
Fractions nested inside fractions (you know the ones)

But it goes further. You can also describe a table or equation in plain English and Bibby generates the LaTeX. "Give me a 3-column table comparing BERT, GPT-2, and T5 on parameters, training data, and benchmark scores" → done. No more fighting with tabular column specifications.

🔍 Feature 4: AI Paper Reviewer — Peer Review Before the Peer Review

One of Bibby's most unique features, and honestly one of the most valuable things I've used as a researcher.

Upload your manuscript PDF and Bibby gives you structured feedback modeled after real peer reviews — with venue-specific criteria for NeurIPS, ICML, ICLR, CVPR, IEEE, ACM, and more.

The feedback isn't generic. It evaluates:

Novelty — is your contribution actually new?
Technical quality — are your proofs/experiments solid?
Clarity — is this readable for someone outside your subfield?
Presentation — figures, tables, formatting
Impact — will this matter to the community?

And it gives you actionable recommendations, not just scores.

I submitted a workshop paper last quarter and ran it through Bibby's reviewer first. It caught a weakness in my experimental setup that I'd been handwaving — the kind of thing a real reviewer would've written a scathing paragraph about. I fixed it. The paper got in.

Try AI Paper Review →

✍️ Feature 5: AI Abstract Generator

Writing abstracts is deceptively hard. You're compressing months of work into 150-250 words while making it compelling, accurate, and keyword-optimized for your target venue.

Bibby's Abstract Generator takes your full paper content (or a PDF upload) and generates a publication-ready abstract that hits the standard structure: motivation, problem, approach, results, impact. You can then refine it with follow-up prompts — "make it less technical," "shorten by 30 words," "emphasize the dataset contribution more."

It's not a replacement for your voice as a researcher. Think of it as a first draft generator that gets you 80% of the way there in 30 seconds instead of 45 minutes.

📖 Feature 6: AI Literature Review Generator

This one is wild.

You describe your research topic. Bibby drafts a literature review section with real, cited references — grouped by theme and methodology, written in coherent paragraphs, formatted with \cite{} commands, ready for you to edit.

It's not making up papers. It's pulling from actual academic databases and synthesizing them. The output isn't perfect — you'll want to verify every citation and add your own framing — but as a starting scaffold, it saves hours of work.

I used it to bootstrap a grant proposal's background section. What typically takes me a full day of reading and organizing took about 2 hours total (including my editing pass on top of Bibby's draft).

🔬 Feature 7: Deep Research Assistant

Sometimes you need to go deep on a topic you're not fully familiar with. Maybe you're pivoting research directions, exploring an adjacent field, or trying to understand the state-of-the-art before formulating a new hypothesis.

Bibby's Deep Research mode does multi-source academic exploration — reading across papers, identifying gaps in the literature, suggesting research directions and hypotheses, and presenting everything with a full citation trail.

It's like having a well-read research assistant who can synthesize a field overview in minutes. Every claim is traceable. Every citation is real. And it points you toward where the interesting open questions actually are.

⚠️ Feature 8: LaTeX Error Detection & Auto-Fix

Remember the 11:47 PM nightmare I opened with? This feature would have ended that in seconds.

Bibby detects compilation errors in real time as you type, explains them in plain English ("You have a missing closing brace on line 47 — here's what the corrected code looks like"), and offers one-click fixes.

It handles:

Missing or conflicting packages
Mismatched braces and environments
Undefined references and citations
Font encoding issues
Multi-file project errors

No more Googling cryptic TeX error messages at midnight. No more Stack Exchange rabbit holes. Bibby just tells you what's wrong and fixes it.

🤝 Feature 9: Real-Time Collaboration (Without the Chaos)

If you've ever tried to co-author a paper over email with 4 collaborators, you know what "merge conflict hell" feels like. Version 7_FINAL_v3_JohnEdits_ACTUALLYFINAL.tex is a horror story we've all lived.

Bibby's real-time collaboration is Google Docs-style live editing but for LaTeX, with:

Unlimited collaborators on Pro plans
In-line comments so you can discuss changes without breaking the document
AI-resolved merge conflicts — when two people edit the same section, Bibby handles it
Version control so you can time-travel through every edit

My last paper had 3 co-authors across 2 time zones. We used Bibby the whole way. Zero version conflicts. First time that's happened in 6 years of collaborative research.

📁 Feature 10: GitHub Sync

For researchers who live in Git (or whose institutions require it), this is a dream feature.

Two-way sync with any GitHub repository — public, private, or institutional. Push directly from Bibby. Pull updates back. Branch support for parallel drafts. Automatic commit messages generated from your edit descriptions.

Your paper is always versioned, always backed up, always accessible from any machine. And your collaborators who prefer to edit locally in VS Code or Vim can still do that — they're just syncing to the same repo.

⏱️ Feature 11: Version History — Time-Travel Through Your Work

Every edit session is automatically saved. Bibby keeps a complete visual diff history — you can see exactly what changed between any two versions and restore with one click.

The named snapshots feature is my favorite part: before a major rewrite, I create a snapshot called something like "Before restructuring intro — NeurIPS draft." Three days later when I've convinced myself the rewrite made things worse, I'm back in 30 seconds.

📄 Feature 12: 5,000+ Journal-Ready Templates

Bibby ships with 5,000+ publisher-approved templates covering IEEE, Nature, APA, ACM, NeurIPS, ICML, Cambridge thesis formats, and hundreds more. Every margin, every font size, every column layout — already configured and ready to go.

Stop spending 45 minutes trying to get the IEEE two-column format right from scratch. Open the template, start writing.

🆓 Bonus: Free Tools for Everyone

Bibby also offers genuinely useful free tools with no account required — because they believe in open science:

LaTeX to Word converter
LaTeX Table Generator
Citation Generator
LaTeX to HTML
PDF to Abstract
LaTeX Word Counter
Common LaTeX Symbols reference
Conference Deadlines tracker

These are legitimately useful standalone tools, not crippled demos designed to upsell you.

What's the Catch? (Honest Pricing Take)

Free plan: Unlimited projects, limited compiles, 2 AI completions/month, 10+ basic templates. No collaboration. Good for trying it out, not enough for serious research.

Pro plan: $9/month (currently 50% off from $18). Unlimited everything — projects, collaborators, AI assist, templates, citations. Version control, Git sync, equation from image, deep research, paper review, priority support, offline mode, export in all formats. 30-day money-back guarantee.

For context: I used to spend ~6 hours per paper on formatting and citation management alone. At even a modest valuation of my time, Bibby Pro pays for itself in the first paper. It's not a close calculation.

Institution plan: Custom pricing for research teams and universities, with SSO, centralized billing, analytics, API access, and on-premise deployment options.

The Thing That Surprised Me Most: Privacy

I was skeptical about an AI tool that sits inside my unpublished manuscripts. Bibby's answer:

Your research data is never used to train AI models. Ever.

They're explicit about this: institutional-grade privacy by design. You can enable or disable any AI feature. Your documents are yours. For anyone working on pre-publication research, this matters enormously — and it's not a throwaway privacy policy line, it's a core product commitment.

My Honest Assessment

Here's what I keep coming back to: Bibby was built by people who understand what research actually feels like. Every feature addresses a real pain point. The AI isn't bolted on for marketing — it's woven into the core workflow.

I went from dreading the "writing phase" of research (mostly because it meant fighting LaTeX for hours) to actually looking forward to it. That's not a small thing.

If you're a researcher who uses LaTeX — or who should be using LaTeX but keeps avoiding it because of the overhead — I'd genuinely encourage you to give Bibby a try. The free tier will let you feel the editor. Spring for a month of Pro and actually use the AI features before deciding.

Start writing free at trybibby.com →

What's your biggest LaTeX pain point? Drop it in the comments — curious if others have found creative workarounds, or if Bibby handles it in a way I haven't explored yet.

Tags: #latex #research #productivity #ai #academicwriting #phd #tools

Creating your Rust Functions in Node.js Using SSVM and docker

Nilesh Arnaiya — Fri, 31 Jul 2020 04:41:12 +0000

Head over to https://github.com/second-state/ssvm-nodejs-starter to get started.

Use Docker:
$ docker pull secondstate/ssvm-nodejs-starter:v1 $ docker run -p 3000:3000 --rm -it -v $(pwd):/app secondstate/ssvm-nodejs-starter:v1 (docker) # cd /app (docker) # ssvmup build (docker) # node node/app.js

Copy the functions example - https://github.com/second-state/wasm-learning/tree/master/nodejs/functions

Replace the files and run your node app.

If you want to try out Machine learning with Rust - https://github.com/second-state/wasm-learning/tree/master/nodejs/ml

Tensorflow Models on Android - Absolute Beginner

Nilesh Arnaiya — Thu, 21 May 2020 17:20:56 +0000

What is Tensorflow Lite?

TensorFlow Lite is an open-source framework created by Tensorflow and optimized to run inference on mobile devices📱 as well as edge devices. TFLite is used for deploying pre-trained models on android, iOS and even on your latest Raspberry Pi

This article is an introduction to TensorFlow Lite and you will learn how to train the model, convert the model, and run inference on your mobile devices.

Requirements

Android Studio

Let's Create a Face Mask 😷 Classifier right in the browser using Google's Teachable Machine

All thanks to Prajna Bhandary for the dataset - Download

Let's Train

Let's use the dataset and upload it to the Teachable Machine

You can add as many classes with as many images you want.
More the Images, More the training time.
Click the Train Model and wait until it completes the training.
The Advanced Tab
- This gives you more freedom to select the basic hyperparameters that help the Neural network learn.
- Epochs - 1 epoch is one run of each of the samples in our data and you want to keep increasing until it you get the desired predictions.
- Batch Size - Divides your samples by the Batch Size you choose and that becomes your batch. For Eg: 112 samples/ 16 = 7 Batches
- 7 batches to be processed by the model to finish 1 Epoch
- Learning Rate - Size of a step that a model makes for choosing the best possible weights in a neural network. The smaller steps take a lot of time to learn which weights would be best and a high learning rate would mean we might miss some best weights possible. So tweak carefully.
Export the model
- Choose a quantized model for reducing the model size and accelerating operations on mobile devices while suffering a very minimal loss.
Deploying on Android Devices
- Let's create a labels.txt file with all our classes and each class on a separate line
- We will use this MobileNet Repository for our Mask Detector and update the files respectively.
- Download or clone the repo and open the project in Android Studio
- Add the labels.txt file in src/main/assets/ Folder.

Replace the model in the assets/ folder with our trained model.
- Change the MODEL_PATH in MainActivity.java
private static final String MODEL_PATH = "mobilenet_quant_v1_224.tflite";
- Connect the USB and RUN the app

Enjoy👏 in thee glory of creating thee Face Mask Detection Model and running it on mobile devices📱.

Additional Resources

If you are using a Custom trained model, use this for conversion Udacity Tflite conversion
If you are using Keras - use this Keras .pb to .tflite
Also, Check out the Udacity Course - Intro to Tensorflow Lite
A Project built on top of TFLite - Smart Mirror