aidedx is a free AI lab assistant for proton and heavy-ion physics. Ask about range, kinetic energy, or stopping power — just speak or type your question in plain language and get a real answer, computed entirely on your own machine, no internet needed once it's set up.
Questions like "how far does a 5 MeV alpha particle go in air?" or "what's the stopping power of 200 MeV/nucl carbon ions in water?" normally mean digging through a form or a data table. aidedx lets you simply ask:
🎤 "What is the range of 40 MeV protons in PMMA?"
→ "The CSDA range of 40 MeV protons in PMMA is 1.285 cm (PSTAR)."
The key point: the answer comes from trusted reference data — NIST (PSTAR/ASTAR), ICRU, and Bethe–Bloch calculations — not from an AI making numbers up. The AI only listens to your question and works out what you're asking; the physics is then computed (via the libdedx library). It's a free web app — no registration, no install, nothing you type or say ever leaves your device.
- Open the app at https://aptg.github.io/aidedx/ — nothing to install.
- Allow the one-time download when prompted. aidedx fetches its AI models (a few hundred MB) and caches them in your browser, so this happens only on your first visit.
- Ask your question — press 🎤 and speak, or type it. For example: "stopping power of 150 MeV protons in water".
- Read the answer. Any assumptions aidedx made (isotope, energy interpretation, program) are noted alongside it.
It's just a web app: no repo to clone, no app to install, no login needed.
aidedx runs in two phases. The first time you visit, it downloads its AI models once and caches them in your browser. After that, every question is handled entirely on your machine, in four steps that each feed the next — speech/text in, then:
Speech recognition is tuned to physics vocabulary rather than everyday speech, and retries if a decode ever fails outright. See the technical documentation for the specifics.
Most AI tools send your words to a company's servers. aidedx does the opposite — the models run locally, in your browser:
The only requirement is that one-time download — fetching the AI's "brain" the first time you use it, while you have a normal connection. After that, your laptop or phone needs no internet at all: it keeps working inside a heavily shielded irradiation room, on a flight to a conference, or anywhere else with no signal.
Early-stage prototype under active development:
- ✅ Works: typed or spoken question → understood → computed → answered, with any assumptions clearly shown so you can correct them by hand; the one-time model download asks first and shows a progress panel; answers link out to dedx_web for full plots.
- 🚧 In progress: Polish-language support (#63) — the language understanding and a Polish voice pipeline exist behind the scenes, but you can't yet switch the app itself into Polish; smarter correction for phrasings the built-in rules miss.
- 🐢 Slow / open: on a computer without a dedicated graphics card, answers can take a few seconds instead of feeling instant — faster models for those machines are still being evaluated (#122).
- 🔭 Planned: spoken answers, wider phrasing coverage, and apps beyond the browser — an Android app is being scoped first (#120), with an installable app for other phones/computers (no app-store download needed) or a full desktop app to follow, reusing the same on-device technology.
- 📖 User guide — coming soon.
- 🛠 Technical documentation — stack, dev workflow, internals, and pointers to every deep-dive doc.
- 🧮 APTG/libdedx — the library computing every number, built on NIST PSTAR/ASTAR and ICRU data.
- 🌐 APTG/dedx_web — the form-driven web tool aidedx complements and reuses.
GPL-3.0-or-later (LICENSE) — see third-party licenses.
Happily vibe-coded with Claude by grzanka.