Skip to content

ResearchTools โ€” find what's wrong before anyone else does

ResearchTools

An AI-assisted toolbox for researcher-professors and graduate students. It finds and fixes what's wrong before a reviewer, a thesis committee, or a grant panel does.

๐Ÿ“š Academic writing

Literature reviews, paper and thesis audits, BibTeX cleanup, reviewer responses, submission packages โ€” a ScholarEval score before the reviewer ever sees it. The figures of a paper or a thesis (TikZ, draw.io, SVG, bitmap) are reviewed with one shared figure-review skill, and /work-ugs-audit grades a folder of undergraduate PDF reports against your rubric with local models only, so a confidential report never reaches a cloud model.

Skills โ†’ ยท Commands โ†’

๐Ÿ› ๏ธ Software review, hardware to come

Code design with a bounded review loop, and the rt-observe dashboard to watch the agents work and log what they did - and to ask it a question out loud: hold P, and a local model answers from the toolkit's own memory, spoken back, nothing leaving the machine. The PCB and 3D CAD chapters are placeholders.

See it in action โ†’ ยท PCB โ†’ ยท 3D CAD โ†’

โš™๏ธ Local-model dev loop

local-writer / local-coder push routine generation to a local Ollama model instead of the cloud โ€” a design/plan/code/review loop that keeps token cost down. Which model fits your GPU is measured, not guessed: opt-local-vram-llm tunes the LLM, opt-local-stt-vram scores speech-to-text models beside it.

Agents โ†’ ยท Local coding and the cloud โ†’

Vibe Design, book cover

The book: Vibe Design

Vibe Design, by Martin J.-D. Otis (in French). Vibe coding, vibe PCB, vibe 3D: a new way of designing complete systems, in which the engineer works in dialogue with coding agents.

Get the book on Ko-fi

Why ResearchTools

Without With ResearchTools
Catch a missing reference or a broken hypothesis flow after the reviewer does An auditor scores the manuscript first, against the same ScholarEval rubric a committee uses
Re-read a cited paper's abstract and hope the citation says what you think it says extract-contributions checks the citing sentence against the paper's own stated contribution
Receive a thesis, an article or a talk as a PDF with no LaTeX source and re-type it before it can be audited pdf2md converts it with mineru to markdown (the UQAC thesis layout, or one document for an article or a talk) and measures what the conversion lost against a second reader of the PDF
Burn cloud tokens on routine docstrings and refactors local-writer / local-coder push that generation to a local model, for free
Guess which local model fits your GPU opt-local-vram-llm and opt-local-stt-vram measure it on the card, beside what already runs there, and print one table to choose from

This page is the homepage of the themed MkDocs Material site (mkdocs.yml at the repo root, brand palette in assets/extra.css โ€” navy #10243E, teal #1F9E8F, amber #C9762F, taken from the project's own design system in post-media/Series Template.dc.html). Nothing here duplicates the manual; every link points at the same file a Claude Code session or a plain GitHub browse would read.

Preview locally: .venv-docs\Scripts\python.exe -m mkdocs serve (the environment comes from .\setup.ps1 -InstallPython), then open the printed http://127.0.0.1:8000. Publish: mkdocs gh-deploy, run by hand whenever the docs change โ€” no GitHub Actions workflow, matching this repo's no-CI/CD policy โ€” then enable Settings โ†’ Pages โ†’ Deploy from a branch โ†’ gh-pages. Links that reach outside docs/ (to README.md, Architecture.md, .claude/CLAUDE.md) stay relative in the Markdown, so they work on GitHub and in the editor. MkDocs only serves docs/, so the build hook scripts/docs/mkdocs_hooks.py sends each such link to its page on GitHub (blob or tree on the branch named by extra.docs_source_ref in mkdocs.yml, currently main), which means those links only resolve once the change is on that branch. Heading ids use GitHub's slug algorithm (toc.slugify in mkdocs.yml), so the double-dash anchors of 06-skills.md jump correctly (a repeated heading still gets MkDocs's _1 suffix where GitHub writes -1). mkdocs build --strict is clean, and fails on a dead link or a dead anchor.

Start here

The manual lives in manual/, one chapter per topic:

# Chapter
00 Purpose & overview
01 Installation
02 Profiles & environment
03 Token management
04 Commands
05 Agents
06 Skills
07 Hooks and guards
08 Rules and conventions
09 Harnesses and mirrors
10 rt-observe & dashboard
11 Local coding and the cloud
12 PCB design, mixed signals
13 3D CAD design, CNC, 3D printing
14 Aider nightly pipeline
15 Security audit and guards
16 Maintenance, checks and CI
17 ThesisTracker integration
18 File locations

Deep-dive references

Longer, single-topic documents that a manual chapter points into rather than repeats:

  • aider-setup.md โ€” the aider nightly pipeline, full manual
  • rt-observe.md โ€” the toolkit-state dashboard, full reference
  • authoring-and-mirrors.md โ€” adding or editing an agent, skill, or command, and regenerating the per-tool mirrors
  • contributor-notes.md โ€” shared conventions (English-only definition files, agent/skill layout, local-model routing, git/GitHub workflow)
  • github-repo-setup-playbook.md โ€” portable step-by-step for bringing another repo's GitHub presentation up to this same standard (README structure, doc-site setup, community files, repo settings, Issue/board/PR process rules), written for a Claude Code session to execute, not for a human to read once

Elsewhere in the repository