Netflix-Style Platform & Recommender - 30/07/2026 07:38 EDT
Budget: ₹12,500 – ₹37,500 INR
I’m building a full-stack movie streaming site that feels as intuitive as Netflix, yet the real heart of the product is a rock-solid recommendation engine. I already have a structured set of user ratings and reviews that can be used to train a collaborative-filtering model; no other data sources are currently available, so the algorithm must squeeze maximum insight from this dataset alone.
Here’s what I need from you:
• A collaborative-filtering engine (memory- or model-based—convince me which makes more sense for scale) that ingests the ratings/reviews and outputs personalised movie suggestions in real time.
• Clean API endpoints so the front-end can request “Top Picks” and “Because You Watched…” rows with millisecond-level latency.
• Core streaming platform features—catalog browsing, playback, basic user profiles—implemented just far enough to demonstrate the recommender in a realistic environment.
Acceptance criteria
1. Given a cold-start user with fewer than five ratings, the engine must still return at least ten distinct titles with measurable diversity (genre spread ≥ 0.4 using Shannon entropy).
2. P95 response time for recommendation calls must remain under 150 ms for a catalogue of 10 k titles and 100 k users.
3. All code delivered in a Git repository with step-by-step setup instructions and a short read-me explaining model choice, hyper-parameters, and any libraries used (Python, TensorFlow, Surprise, etc.).
If you’ve previously shipped a recommender on similar data, I’d love to see the demo or metrics.
Here’s what I need from you:
• A collaborative-filtering engine (memory- or model-based—convince me which makes more sense for scale) that ingests the ratings/reviews and outputs personalised movie suggestions in real time.
• Clean API endpoints so the front-end can request “Top Picks” and “Because You Watched…” rows with millisecond-level latency.
• Core streaming platform features—catalog browsing, playback, basic user profiles—implemented just far enough to demonstrate the recommender in a realistic environment.
Acceptance criteria
1. Given a cold-start user with fewer than five ratings, the engine must still return at least ten distinct titles with measurable diversity (genre spread ≥ 0.4 using Shannon entropy).
2. P95 response time for recommendation calls must remain under 150 ms for a catalogue of 10 k titles and 100 k users.
3. All code delivered in a Git repository with step-by-step setup instructions and a short read-me explaining model choice, hyper-parameters, and any libraries used (Python, TensorFlow, Surprise, etc.).
If you’ve previously shipped a recommender on similar data, I’d love to see the demo or metrics.