AI/ML Engineering Leadership
Thirteen years of building machine learning products and the organizations that ship them.
Alok Upadhyay is an AI/ML engineering leader based in Seattle: currently a Senior Engineering Manager at Apple, leading initiatives across Foundation Model Post-Training, Recommendations, Search, and AI Safety, delivered on-device and in the cloud.
He spent the previous twelve years at Amazon, where he was a founding engineer three times over: Amazon Home Services (2014), Alexa Identity (2017), and the Ambient Recognition & Authentication group (2018): before moving into engineering management and, ultimately, manager-of-managers roles across Amazon AGI and Prime Video, leading organizations of roughly twenty engineers, applied scientists, and technical program managers.
What makes his profile unusual is the combination: he runs ML organizations and keeps shipping first-hand technical work. He is an inventor on 4 granted US patents: cited as prior art in subsequent filings by engineers at Google and Microsoft: author of three sole-author papers at ICLR 2026 workshops and MathAI 2026 (oral), recipient of a Best Reviewer Award at an ICLR 2026 workshop, and co-author of Amazon's Authentication Confidence Levels standard, the framework governing identity-verification confidence across all Alexa devices and third-party developers.
How He Leads
Science to production
Research that never ships is a preprint; systems without science plateau. The job of an ML org is the loop between the two: eval-gated launches, online learning, and ground truth as a first-class system.
Zero-to-one under ambiguity
Five flagship identity launches at Alexa: Voice ID through Recognition Moments: were built without an established playbook, coordinating 23+ external teams and substantially growing the base of personalizable devices.
Grow the builders
Organizations compound through people: 15+ promotions driven across teams, including ICs grown to Staff Engineer and engineers developed into managers: with low attrition through the heaviest workloads.
Fly the airplane
As an FAA-certificated private pilot, he brings cockpit discipline to engineering: pre-committed decision criteria, graceful degradation, and never trading away safety margin for speed.
The Post-Training Era
The center of gravity in applied ML has shifted from pre-training models to post-training them, and to the evaluation and safety discipline that lets them ship.
Alok's current work sits at that center: supervised fine-tuning, reinforcement fine-tuning, and GRPO for production LLM experiences; recommendation and search systems in the LLM era; and AI safety as an engineering discipline: treating evals the way high-consequence systems treat launch reviews. His 2026 research asks the same questions from the outside: whether VLM identity judgments are logically consistent, and whether LLM recommenders obey the preference axioms classical systems were built on.