active · Architected & built ·open-source

↻ dhilipsiva.dev

RustWASMZolaLoRA

Problem

A personal site is usually a brochure. I wanted one that demonstrates the thesis instead of stating it — that the gap between fluent and true is real, and that the honest way to ship AI is on-device, inspectable, and open.

Approach

A Zola static site on the QUINE design system, zero backend. The AI stack runs entirely in your browser: candle (Rust compiled to WebAssembly) drives two LoRA fine-tunes of me at /chat — a SmolLM2-135M and a Qwen2.5-0.5B that opens the site's apps itself — plus Whisper voice input. The real nibli engine runs the Transparency Triad at /nibli, proof trees and all.

Outcome

You're looking at it. A quine of a website: the repo contains the instructions for building itself, including finetune/ — the follow-along runbook that turns a fact sheet and some seed answers into the model impersonating me. Everything is open source, training data included.