WriteTank

A privacy-first AI writing assistant for LaTeX to easily write anything, and only lives on your own computer.

Comment

Inspiration

As a PhD researcher, I spend countless hours writing in LaTeX on Overleaf. Writing academic papers, theses, scientific reports, CV, research proposals and personal statements is time-consuming and mentally draining. I wanted an AI assistant that:

  • Understands LaTeX formatting and syntax.
  • Works directly inside my editor.
  • Helps structure, refine, and summarise writing, and constantly giving me comments and suggestions for improvement.
  • Keeps my data and idea private.

Most existing assistants send sensitive documents and personal information to the cloud, which is not acceptable or convenient for research or confidential work, for security reason. That inspired me to build WriteTank.

What it does

WriteTank is a privacy-first academic writing assistant powered by the gpt-oss:20b model. It integrates into Overleaf through a Chrome extension, offering two main features:

  • Q&A Mode: highlight selected text or automatically scan for visible text in the editor, aligning to user's perspective, ask a question, and get LaTeX-formatted answers.
  • Auto-Coach Mode: automatically suggests structure, details to add, and a polished paragraph, with TODOs for missing citations. If the text is well-written, it will also give encouraging comments.

All model inference runs locally on-device via Ollama, ensuring complete privacy.

How I built it

  • Frontend: Chrome extension built with TypeScript, Vite, and CRXJS.
  • Integration: Content scripts inject WriteTank panels into Overleaf for Q&A and Auto-Coach.
  • Backend: Model calls handled by Ollama at localhost:11434.
  • Model: gpt-oss:20b, tuned with prompt engineering, stop sequences, and token limits, to ensure smooth user experience.
  • Output: Strict LaTeX formatting for seamless integration with academic and personal documents.

Challenges I ran into

  • Running a 20B model locally on a laptop required trimming inputs, capping outputs, and tuning Ollama options.
  • Overcoming Chrome extension security policies (CSP, CORS) to allow local API access.
  • Designing a minimal, non-intrusive UI inside Overleaf that feels natural. Should be self-explanatory, easy to use and follow around.
  • Managing time constraints — ideally, this would work with any text editor, but I focused on Overleaf for the hackathon, and I have developed this particularly with a focus for academic and personal writing, although this can be easily extend to writing any materials, in any languages.

Accomplishments that I am proud of

  • Built a smoothly working Chrome extension that connects Overleaf to a local LLM.
  • Successfully integrated LaTeX-aware prompts for structured academic output.
  • Demonstrated that powerful AI writing assistance can run fully offline.
  • Created a clean, modern UI that feels natural for researchers.

What I learned

  • How to run and optimise large local models with Ollama.
  • Practical prompt engineering for LaTeX formatting and context awareness.
  • Building and packaging Chrome extensions with Vite + CRXJS.
  • Trade-offs between different quantisation methods (MXFP4 vs Q4_K_M) on Apple Silicon.
  • The importance of perceived speed — streaming and partial answers can feel faster than raw throughput.

What's next for WriteTank

  • Extend beyond Overleaf to support any text editor.
  • Add real-time streaming responses for faster user experience.
  • Build domain-specific fine-tuned models for specialised academic fields.
  • Position WriteTank as a general-purpose local AI agent for confidential or sensitive writing, from research papers to business reports.

Built With

Share this project:

Updates