Devansh Tandon

Devansh Tandon is a principal product manager at Meta developing foundation models and large-scale recommendation systems. At Google and YouTube, he helped adapt Gemini for video recommendations, turning enormous content catalogs into a machine-readable vocabulary for generative retrieval.

Tandon studied computer science and economics at Yale, graduating magna cum laude with honors in both disciplines. He entered Google through its associate product manager program and spent seven years working across advertising, Search, Discover, YouTube, and Google DeepMind. His account of leaving Google describes an enduring interest in personalized, AI-powered consumer products; at Meta, his work encompasses recommendation systems associated with Instagram, Facebook, and advertising.

• Semantic IDs: Tandon coauthored research on content-derived item identifiers that replace arbitrary identifiers with structured tokens. At YouTube, video titles, descriptions, transcripts, audio, and frames become multimodal embeddings compressed with RQ-VAE, helping models understand new and infrequently watched content.

• Large Recommender Model: His team adapted Gemini into a YouTube-specific checkpoint that connects natural language with video tokens and learns relationships from viewing sequences. The resulting model can generate recommendation candidates directly; PLUM, which Tandon coauthored, extends this approach into an industrial-scale recommendation framework.

• Offline recommendation generation: Serving foundation models across an enormous, constantly changing video catalog creates significant cost and latency constraints. Tandon described precomputing next-video candidates and serving them through inexpensive lookups, while continued pretraining and semantic representations help incorporate newly uploaded videos.

• User-steerable recommendation feeds: Tandon argues that personalized feeds are an overlooked consumer-AI opportunity. He envisions systems that explain recommendations, respond to natural-language preferences, and eventually combine personalized retrieval with generated content.