10. Interview Prep Guide

2-minute script

"I built a static browser demo that runs IndicTrans2 entirely client-side with ONNX Runtime Web. You pick model size and precision, weights download from Hugging Face into Cache Storage, then a custom encoder–decoder loop with KV cache translates English and Indic languages. Script transliteration keeps non-Devanagari text aligned with the model. There's no server — the companion export repo produces the ONNX artifacts."

Likely questions

Q	A
WebGPU vs WASM?	WebGPU faster when stable; quants often safer on WASM
Why custom decode?	Control over exported graphs, dual tokenizers, FLORES tags
Privacy?	Inference local; still trusts HF/CDN for code and weights
Scale limits?	Browser RAM; 1B fp32 is multi-GB

Weak spots

No tests; main-thread inference can jank UI; no cache purge on unload.

Proud of

Making optimized Indic NMT usable in a pure static page with real metrics.