B5 Studios · 2026 to present
VP of Engineering & AI
I’m building the engineering and AI organization from the ground up. I set the technical direction, hire the team, and stay hands-on as we develop tools for creative production.
Ritvik
Shrivastava
VP of Engineering & AI at B5 Studios
At B5 Studios, I’m building the engineering team and shaping how we use AI in film, television, and new forms of storytelling.
Previously, at Meta, I led work on proactive AI agents for smart glasses. I also worked on Muse and collaborated with Meta Superintelligence Labs and TBD on Health AI and Accessibility AI.
Experience
Meta Reality Labs · 2024 to 2026
Technical Lead, AI Agents for Wearables
I led modeling, data, and evaluation for proactive AI agents on smart glasses. We worked on how to split tasks between on-device models and larger language models, and how to test whether an intervention was timely and useful.
One accessibility application was helping a Deaf or hard-of-hearing wearer notice a sound they wanted to hear. Our audio research asked when to alert the wearer and how to avoid repeating an alert. In a separate project, we studied how an assistant could help someone recover when a multistep task went off plan.
Beyond wearables, I contributed to Muse. My Health AI work included collaborating with Meta Superintelligence Labs on post-training frontier models.
Cisco · 2020 to 2024
Engineering Lead, Machine Learning
I started on MindMeld and Webex Assistant, then led machine learning for Webex Contact Center. One practical question was how to identify why customers were calling across thousands of conversations. Our work compared models on both the quality of those answers and the cost of generating them.
Research
Papers I’ve coauthored on proactive assistants, deployed language models, and the ways people use language online. Google Scholar ↗
9 publications
Plan, Watch, Recover
A Benchmark and Architectures for Proactive Procedural Assistance
A benchmark and dynamic planning approach for assistants that guide multistep tasks and help people recover from mistakes.
Research details
Introduces EgoProactive and Pro²Bench to evaluate intervention timing, task completion, and response quality. The model maintains an explicit plan and revises it after guidance or a deviation, including recovery steps.
I’ll Keep an Ear Out
Teaching AudioLLMs Proactive Audio Assistance
Audio models that monitor a stream from a user’s intent and decide when a relevant sound deserves an alert.
Research details
Interrupt and Silent Modeling puts intervention decisions into model decoding. The work considers event onset, continued relevance, irrelevant audio, and duplicate alerts, motivated by wearable assistance for Deaf and Hard of Hearing users.
LLM-Based Insight Extraction
for Contact Center Analytics and Cost-Efficient Deployment
Extracting the reasons behind customer calls while evaluating the quality and cost of the models used in production.
Research details
Connects call-driver generation to topic modeling, call classification, trend analysis, and FAQ generation. Compares proprietary, open-weight, and fine-tuned models alongside strategies for practical deployment.
“Laughing at you or with you”
The Role of Sarcasm in Shaping the Disagreement Space
Studying how sarcasm changes the way we recognize agreement and disagreement in online conversations.
Research details
Combines sarcasm and argumentative-relation modeling through linguistic features and neural multitask learning. The experiments show that accounting for sarcasm improves recognition of conversational disagreement.
Topical Stance Detection for Twitter
A Two-Phase LSTM Model Using Attention
An attention-based model that separates subjectivity detection from predicting a stance toward a topic.
Research details
T-PAN first distinguishes neutral from subjective posts, then identifies whether a subjective post supports or opposes the topic. This separates topical stance from general sentiment.
Topic Lifecycle on Social Networks
Analyzing the Effects of Semantic Continuity and Social Communities
Following the evolution of topics across time, changing hashtags, and different online communities.
Research details
Uses semantic and temporal relationships to connect hashtags into topics. Analyzes how community boundaries influence the way topics emerge, persist, and change.
EmTaggeR
A Word Embedding Based Novel Method for Hashtag Recommendation on Twitter
Recommending hashtags using the semantic relationship between a post and learned hashtag representations.
Research details
Compares hashtag-specific and global embedding models, then ranks hashtags using semantically similar posts. The approach is designed to remain lightweight and scalable.
SemTagger
A Novel Approach for Semantic Similarity Based Hashtag Recommendation on Twitter
Earlier work on semantic similarity for recommending hashtags in social media.
Research details
Published in the proceedings of the 14th International Conference on Natural Language Processing. The paper and complete citation are available through the ACL Anthology.
A Paraphrase and Semantic Similarity Detection System
for User Generated Short-Text Content on Microblogs
Recognizing paraphrases and semantic similarity in noisy, informal text from microblogs.
Research details
Develops features for an SVM-based system and evaluates them on Twitter paraphrase data and the Microsoft Paraphrase corpus, examining performance across different text settings.
No publications match that search. Try another title, author, or year.
About me
I came to NLP through language itself. I grew up reading and writing prose and poetry, then found myself studying how people disagree, use sarcasm, and say the same thing in different words.
I studied computer science at NSUT and earned my master’s at Columbia. Along the way, I spent time at IBM Research, IIT Delhi, and the Max Planck Society.
When I travel, I like to explore cities on foot, usually with a coffee stop or a bakery in mind. Back in San Francisco, you can usually find me making a case for a dubious answer at pub trivia or hacking on whatever AI model just dropped. Weekends leave room for a Chelsea match, a new recipe, and a movie.