Aditya Gautam is an AI researcher and engineer whose work connects recommendation systems, content integrity, and efficient multimodal models. His career spans embedded engineering and machine learning at Google and Meta, where his work has included understanding user interests and addressing content problems in Reels.
From embedded systems to machine learning
Gautam’s career in engineering began at Qualcomm in India, where he worked on embedded systems from roughly 2011 to 2014 before moving to Carnegie Mellon University. A lecture by Tom Mitchell introduced him to a compelling application of probability: distinguishing spam from legitimate email. He pursued machine-learning coursework and, after graduation, spent about a year working on Zillow’s estimation algorithms.
He then spent three and a half years at Google, including work as a founding engineer for a computer-vision venture within its Area 120 incubator. After a brief startup role, he joined Meta. His work developed from prediction and ranking systems into generative models that interpret content, retrieve supporting information, and coordinate decisions across specialized agents.
His work at Meta included leading generative AI initiatives for Reels involving user-interest exploration and policy understanding, with responsibilities spanning multimodal model training, agentic systems, and production pipelines. These problems require more than recognizing a video’s dominant subject: a brief insertion can change its meaning, while a transformed copy can obscure its original creator. His approach combines granular visual understanding with retrieval and the constraints of processing large volumes of user-generated content.
Evidence, usefulness, and specialized models
Misinformation beyond detection: Gautam’s 2025 misinformation framework proposes five agents responsible for maintaining trusted information, classifying questionable content, retrieving evidence, generating corrections, and verifying the result. The aim is to make corrections traceable to supporting material while allowing components to improve independently. The architecture remains a proposal, with implementation and empirical evaluation left for future work. Its acknowledged limitations include stale information, coordination failures, bias, and the cost of continuous retrieval.
Usefulness beyond relevance: With Mouly Dewan, Jiqun Liu, and Chirag Shah, Gautam co-authored research on evaluating search results by whether they help someone accomplish a task. Their Task-aware Rubric-based Usefulness Evaluation, or TRUE, incorporates search-session context and behavioral signals into model-generated judgments. A document can match a query without advancing the user’s goal; evaluating usefulness requires understanding that goal and the surrounding search session. Their experiments found improvements under several conditions, with uneven results across models and datasets that reinforce the need to evaluate the judging method itself.
Small models matched to the job: Gautam favors specialized vision-language models for narrowly defined internal tasks. His approach to multimodal video systems separates perception, retrieval, and review when those functions require different expertise, while retaining a single-agent approach when it is sufficient. The practical distinction is between recognizing what a video contains and deciding what that content means in context. Detecting a brief, unexpected insertion requires attention to changes within the video; identifying a transformed copy also requires finding related material. Training on actual user-generated content and testing distillation and quantization align model capability with serving cost. General coding ability contributes little to a model whose job is interpreting video.
Evaluation that follows the failure: Gautam treats evaluation as a continuing engineering loop. Production samples pass through an in-house model judge and human review, helping locate failures in retrieval, tool use, and reasoning. Those findings inform further training, including direct preference optimization. He evaluates retrieval quality, latency, reasoning budgets, and token costs alongside task accuracy, and checks whether automated judges drift away from human assessments. Efficiency also begins before inference: compressing similar frames, reusing decisions for repeated viral content, and pruning candidates by metadata can avoid unnecessary processing.
Gautam’s thinking about compute efficiency extends beyond model size. Memory bandwidth, retrieval speed, network latency, and orchestration all affect whether an agent workflow is practical. His emphasis is on specialized models within a measured process that can improve as content and user behavior change.
Aditya Gautam of Meta explains how specialized vision-language models divide perception, retrieval, and review to catch brief content changes and investigate copied videos—and how training, evaluation, and selective processing make that architecture practical at scale.
Time-linked clip evidence makes a brief topic change inspectable: the perceiver extracts it, and the reviewer compares it with surrounding content.
Offline indexing and clustering prepare the corpus for online retrieval; topics, embeddings, and entities supply different routes to similar clips and authors.
Choose specialized model sizes and reasoning budgets against domain quality, latency, and cost, while checking the LLM judge against fresh human labels.
Frame compression, cached decisions, and metadata pruning save work at different stages; ongoing evaluation and monitoring keep those optimizations useful as the data changes.
▶ Watch ↗