Alok UpadhyayWriting on multimodal AI, ML systems at scale, post-training, retrieval and recommendation.https://alok.ai/en-usRLHF Post-Training: Designing the Training and Serving Systemshttps://alok.ai/blog/2026-07-15-what-rlhf-actually-optimizes/https://alok.ai/blog/2026-07-15-what-rlhf-actually-optimizes/RLHF is usually explained as an objective. An explainer on the systems it implies: four model copies resident at once, an inference engine and a trainer in the same job, and a serving path where the thing you optimised against never ships.Wed, 15 Jul 2026 00:00:00 GMTexplainerllmpost-trainingrlhfml-systemssystem-designSafety Policy as Retrieved Contexthttps://alok.ai/blog/2026-06-11-safety-policy-as-retrieved-context/https://alok.ai/blog/2026-06-11-safety-policy-as-retrieved-context/Most safety behaviour is frozen into weights at training time, while the policy it encodes changes weekly. An explainer on retrieving the clauses that apply to a request and inlining them as explicit constraints, and the failure modes that introduces.Thu, 11 Jun 2026 00:00:00 GMTexplainerai-safetyragllmguardrailsDesigning a RAG Search System: Indexing, Retrieval and Groundinghttps://alok.ai/blog/2026-05-14-rag-search-system-design/https://alok.ai/blog/2026-05-14-rag-search-system-design/Retrieval-augmented generation is usually drawn as three boxes. An explainer on the parts that decide whether it works: chunking and reindexing, hybrid retrieval and reranking, permissions enforced at retrieval, and the difference between a bad answer and a bad retrieval.Thu, 14 May 2026 00:00:00 GMTexplainerragsearchretrievalllmsystem-designTransformer Recommenders: Designing the Training and Serving Systemshttps://alok.ai/blog/2026-04-16-transformer-recommenders-system-design/https://alok.ai/blog/2026-04-16-transformer-recommenders-system-design/Sequential transformer recommenders changed what the model is. An explainer on the systems work that decides whether that change survives contact with production: point-in-time correctness, the retrieval/ranking funnel, and the skew that quietly eats the gains.Thu, 16 Apr 2026 00:00:00 GMTexplainerrecommendation-systemstransformersml-systemssystem-designCan Geometry Predict Whether a Face Matches a Voice?https://alok.ai/blog/2026-03-14-riemannian-geometry-multimodal-biometrics/https://alok.ai/blog/2026-03-14-riemannian-geometry-multimodal-biometrics/We show that the intrinsic Riemannian geometry of pretrained neural network embedding spaces predicts how well face and voice can be matched across modalities: without any cross-modal training.Sat, 14 Mar 2026 00:00:00 GMTmultimodal-airiemannian-geometrybiometricscomputer-visionresearchDo LLM Recommenders Obey Preference Axioms?https://alok.ai/blog/2026-02-23-llm-recommenders-preference-axioms/https://alok.ai/blog/2026-02-23-llm-recommenders-preference-axioms/We test whether LLM-based recommender systems satisfy classical rationality axioms from social choice theory, and find that all models violate every axiom, with a striking dichotomy between pairwise and set-based reasoning.Mon, 23 Feb 2026 00:00:00 GMTlarge-language-modelsrecommender-systemssocial-choice-theorylogical-reasoningresearchAre VLM Identity Judgments Logically Consistent?https://alok.ai/blog/2026-02-19-vlm-identity-logical-consistency/https://alok.ai/blog/2026-02-19-vlm-identity-logical-consistency/We test whether vision-language models obey symmetry and transitivity when judging if two images show the same person, and find a striking accuracy-consistency trade-off.Thu, 19 Feb 2026 00:00:00 GMTvision-language-modelsperson-re-identificationlogical-reasoningcomputer-visionresearchBuilding Multimodal AI Systems That Serve a Billion Recognitions a Dayhttps://alok.ai/blog/2025-05-06-building-multimodal-ai-at-scale/https://alok.ai/blog/2025-05-06-building-multimodal-ai-at-scale/Lessons learned from building and scaling multimodal person recognition at Amazon: fusing voice, face, Bluetooth, and behavioral signals to identify users across millions of devices.Tue, 06 May 2025 00:00:00 GMTmultimodal-aiperson-recognitiondistributed-systemsmachine-learning