The fine print,
in plain sight.
How this site is built, what it runs, what it stores and what it does not. No analytics, no cookies, no chat persistence.
The machine
A real 0.5B language model, loaded on demand, running on the visitor's GPU. The conversation is grounded in fact cards about Parth's work and is never sent to a server — unless you turn on Web mode, an opt-in toggle that sends only a general question to Wikipedia.
- Model
- Qwen2.5-0.5B-Instruct-q4f16_1-MLC
- Download
- ~265 MB · cached after the first load
- GPU memory
- ~945 MB
- Context
- 4,096 tokens
- Runtime
- WebLLM · WebGPU · Web Worker
- Licence
- Apache 2.0
Plain code reads each question first — which project, which detail, whether it follows on from the last answer — and hands the model the two or three fact cards that answer it. The model writes the reply; a sentence carrying a number that isn't in those facts is dropped, and the sources beside each answer are the check. General questions are answered from the model's own knowledge, or from Wikipedia with Web on. Without WebGPU, the site answers instantly from the same fact cards.
Privacy
- No analytics, no cookies, no tracking pixels, no third-party scripts.
- No chat persistence. Reload the tab and the conversation is gone.
- No contact form and no email address harvesting.
- Fonts are self-hosted through Next.js; no font CDN request.
- Web mode is off by default. When it is on, only the search question travels to Wikipedia — the conversation itself never does.
- The only requests beyond the page are the explicit model download and, with Web mode on, the Wikipedia lookup you asked for.
- Model weights are cached by your browser; the conversation is not.
Type
Colour
Stack
- Framework
- Next.js 16 · React 19 · TypeScript
- Styling
- Tailwind CSS 4 · hand-written design tokens
- Type
- Geist Sans · Geist Pixel Square · Geist Mono
- Motion
- CSS transitions and keyframes, reduced-motion aware
- Model
- WebLLM 0.2.87 · WebGPU · Web Worker
- Understanding
- Subject and detail routing over ~80 fact cards · MiniSearch 7 for keyword lookups
- Icons
- Lucide
- Hosting
- Vercel · static output · canonical parthvarekar.in
Accessibility
- Semantic landmarks, a skip link and a visible focus ring.
- Keyboard-operable navigation, filters, tabs, chat and command palette.
- Reduced-motion support: animation and smooth scrolling are disabled.
- Colour pairs chosen for readable contrast on the periwinkle canvas.
- Live regions announce chat streaming and filter counts.
- Images carry descriptive alt text; decorative motifs are hidden.
Credits
- GeistVercel
- WebLLMMLC AI
- Qwen2.5 0.5BHugging Face
- Next.jsVercel
- LucideLucide contributors
- MiniSearchLuca Ongaro
The projects, screenshots, writing and design are Parth Varekar's own work. The site is built in Mumbai. The canonical address is parthvarekar.in.