Embedding-based retrieval powers recommendation and discovery features across modern news platforms, yet serving dense float vectors at scale introduces substantial latency and memory costs. We present a two-stage retrieval pipeline deployed at NDTV, one of India's largest news organizations, that replaces exhaustive float search with a fast binary shortlist followed by a lightweight float rerank. At the core of our approach is a learned binary projection trained with a contrastive objective to preserve the neighbor structure of the original embedding space. We evaluate the pipeline on a corpus of 47,300 news articles across three embedding models (Gemma, Nomic, Qwen) and two production-motivated retrieval scenarios: a top-20 recommendation task requiring \geq 98% recall, and a top-5 precision-critical task demanding near-perfect accuracy. Our results show that even moderate bit-widths (1024 bits) recover ≥ 95% of float-exact neighbors, while our recommended operating point of 3072 bits achieves 98% top-20 recall through a 50-candidate shortlist and near-perfect (≥ 99.9%) top-5 recall through a 200-candidate shortlist. Notably, RaBitQ-style scalar corrections---widely adopted in the literature---consistently failed to improve and often degraded retrieval quality for both learned and raw binary codes. At million-scale (1M documents), the binary shortlist delivers 4--6× speedup at 3072 bits and up to 70× at lower bit-widths on a single 16-vCPU cloud instance with no GPU, reducing per-batch latency from 2.5--3.5s to under 600ms. Prior to deployment, an editorial acceptance evaluation by five senior NDTV editors confirmed that the binary pipeline's recommendations are indistinguishable from float-exact results across diverse news topics. We further introduce rotation-aware binary training, a new training scheme that bakes a random orthogonal transform into the learned projection, consistently improving recall at higher bit-widths while narrowing the train--eval generalization gap. These findings generalize across all three embedding families with no model-specific tuning.