I’m a Principal Research Scientist at Meta Wearables, leading foundational AI research and 0→1 product programs. My current work spans Omni model architecture and pretraining, recursive self-improvement, and real-time AI.
Model development, systems architecture, and launch.
Current research · Meta Wearables
I lead architecture and pretraining for models with speech, vision, and text inputs, and speech and text outputs. The work connects multimodal learning, context, and natural interaction in a single model.
Llama Mini at Meta
I led the Llama Mini foundation-model program at Meta, connecting architecture and pretraining with capability transfer and efficient deployment. The released research includes a 1.4T-token initial pretraining stage, simulation-guided data mixtures, specialist checkpoint merging, and quantization-aware training.
I also built on-device foundation models powering Ray-Ban Meta, jointly optimizing model quality, inference latency, memory footprint, and power consumption.
Program lead · Joint-first, joint-last and corresponding author of the released model report.
Assistant systems · Meta
I was the Engineering DRI for Meta AI on Ray-Ban Meta from concept through launch, spanning AI interaction design, model training, and execution across glasses, phone, and cloud.
I defined AI interaction capabilities for the wearable assistant and built device–server orchestration connecting on-device invocation and low-latency local execution with server routing for complex queries.
Ray-Ban Meta commercial
Defining how people invoke, converse with, and act through a wearable assistant.
The commercial featuring Chris Pratt, Chris Hemsworth, and Kris Jenner illustrates the voice and visual AI interaction experience.
Watch the Chris Pratt commercialRay-Ban Meta · “Hey Meta, Who Eats Art?”
Multimodal architecture and pretraining across speech, vision, and text: aligning asynchronous streams and retaining useful context for natural interaction.
Related language-model research investigates how positional information transfers through distillation during context extension. Research on context transfer
Current research into how AI systems can improve their own capabilities through iterative learning, feedback, and evaluation.
Architecture, training recipes, distillation, and hardware-aware design that bring model capability within practical compute, memory, latency, and power budgets. Llama Mini at Meta is one example of this work.
Co-authored research studies compact sparse-expert architectures, weight-decomposed experts, and training objectives for efficient expert offloading. Research on sparse-expert architectures
Streaming perception, human intent, and interaction timing: connecting multimodal models with responsive execution across devices and servers. I’m interested in extending these foundations to embodied agents that interpret changing sensory context and coordinate timely decisions under compute and connectivity constraints.
My co-authored work trains speech models to issue tool queries while the user is still speaking. Research on learned interaction timing
Research leadership, engineering management, and senior technical lead roles.
I have led virtual research teams and tiger teams of fewer than 10 people, as well as cross-functional programs involving 70+ people. Across engineering organizations and external partners, I set technical direction and bring teams together to advance foundation models and ship AI systems.
2020–present
Principal Research Scientist
Previously held engineering management and senior technical lead management roles. Engineering DRI for Meta AI on Ray-Ban Meta from concept through launch; provided technical direction across four engineering organizations spanning training, runtime, and product.
2016–2020
Founding Research Engineer, Natural Language Understanding
Co-designed Siri’s first shipped deep-learning NLU models, replacing SVM/CRF systems with BiLSTM classifiers and parsers. Contributed to contextual dialogue and robust speech-language integration.
Ph.D. · 2017
Computational intelligence and complex systems
Research on computational modeling and optimization of complex dynamical systems, with applications in systems biology. Developed DOPS, a hybrid of multi-swarm search and adaptive optimization.
DOPS paperAdithya Sagar is a Principal Research Scientist at Meta Wearables, leading foundational AI research and 0→1 product programs. His current work spans Omni model architecture and pretraining, recursive self-improvement, and real-time AI. He has led small language-model research, including Llama Mini, built on-device foundation models for Ray-Ban Meta, and served as Engineering DRI for the wearable assistant from concept through launch. His systems work includes AI interaction design and device–server orchestration. Previously at Apple, he co-designed Siri’s first shipped deep-learning NLU models. He holds a Ph.D. from Cornell University.
Media coverage
“Intelligence per token, intelligence per word and intelligence per dollar.”
A conversation with Shereen Bhan on efficient models and deploying intelligence on phones and wearables.
Watch the CNBC-TV18 conversationResearch perspective
In my MobileLLM-Pro launch post, I discuss how models that reason locally and collaborate with larger cloud systems connect to the direction of Meta’s Ray-Ban Display and Neural Band launches.
Read the perspectiveFoundation models, embodied intelligence, real-time AI, and research-to-product leadership.