I built a TikTok-style shopping feed with a recommender system — in pure client-side JS

Developer Bhaumik Tandan built ThodaSa, a TikTok-style shopping feed for the Indian market with a recommender system that runs entirely in client-side JavaScript. The static site on GitHub Pages uses 13-dimensional product feature vectors and a weighted engagement profile stored in localStorage to personalize the feed without any backend or tracking servers. The project is open source on GitHub.

TL;DR: I built ThodaSa https://thodasa.com — a reels-style impulse-shopping demo for the Indian market. You scroll products like Instagram reels, and the feed learns your taste — with no backend, no login, and no tracking servers. The whole recommender is ~150 lines of client-side JavaScript. Code is on GitHub https://github.com/Bhaumik-Tandan/thodasa . Quick-commerce apps in India Blinkit, Zepto, Meesho figured out something interesting: shopping is entertainment. I wanted to push that to its logical end — what if the store was literally a reels feed? One product per screen, full-bleed photo, swipe up for the next dopamine hit, everything under ₹499. And because a feed is boring if it's the same for everyone, it needed a recommender system. The catch: this is a static site on GitHub Pages. No servers. So the recommender had to live entirely in the browser. Every product gets embedded as a 13-dimensional feature vector — no ML libraries, just an array: // 0..7 category one-hot snacks, beauty, gadgets, home, ... // 8..10 price bucket one-hot low ≤150, mid 151–300, high 300 // 11 deal flag // 12 highly-rated flag ≥4.5 export const vecOf = p = { const v = new Array 13 .fill 0 v CATS.indexOf p.category = 1 v p.price <= 150 ? 8 : p.price <= 300 ? 9 : 10 = 1 if p.deal v 11 = 1 if p.rating = 4.5 v 12 = 1 return v } The user's taste is a weighted running sum of the vectors they engage with, persisted in localStorage : | Signal | Weight | |---|---| | Purchase | +10 | | Add to cart | +8 | | Wishlist | +5 | | Share | +4 | | Dwell 4s on a card | +2 | | Flick past in < 1.2s | −1 | | Un-wishlist | −3 | Dwell time comes from an IntersectionObserver on the snap-scroll feed — if you pause on a card, that's a signal; if you flick past it instantly, that's a signal too. Every new session decays the profile by 0.85, so recent taste dominates. On each visit, every product is scored: score = cosine profile, vecOf product + Math.random 0.15 // jitter - Math.min seenCount, 5 0.06 // fatigue penalty Then the feed interleaves: two "exploit" cards best matches, badged ✨ For you for every one "explore" card random from the long tail, badged 🎲 Fresh find . Pure exploitation makes an echo chamber; the exploration slots keep the feed a discovery machine. Cold start fewer than 3 signals falls back to a hand-curated launch order. scroll-snap-type: y mandatory + one 100dvh card per product gets you TikTok-feel scrolling with zero JS scroll handlers. IntersectionObserver is a shockingly good implicit-feedback sensor.Scroll a few beauty products and reload — watch the feed rearrange itself. Then check the "Your vibe" widget in the wishlist to see what it learned about you. Roast the code, star the repo, or tell me what you'd impulse-buy under ₹499.