Bhavith Chandra
MS CS @ NYU · Mechanistic Interpretability · World Models · AI Safety
I am a Master's student in Computer Science at New York University, working on world models, mechanistic interpretability, and AI safety. Graduate Research Assistant at the COPHEE Lab under Prof. Vedant Das Swain. Teaching Assistant to Prof. Yann LeCun and Prof. Alfredo Canziani for Deep Learning (DS-GA 1008), Fall 2025 and Fall 2026. This fall I join Frisson Labs as an ML Research member and the Iliad intensive as an AI Safety Research Fellow. This site collects my research, my writing, and notes from Elemental Research Lab.
About Me
I am a Master’s student in Computer Science at New York University working on world models, mechanistic interpretability, and AI safety. This fall I am joining Frisson Labs as an incoming ML Research member (world models applied to voice and gaming), and the Iliad intensive as an incoming AI Safety Research Fellow (agent foundations, modular systems, world models). Over the summer I was an AI Research Engineer at a stealth New York lab focused on enterprise world models and world-model interpretability.
At NYU I am a Graduate Research Assistant in the COPHEE Lab under Prof. Vedant Das Swain. My thesis is on interpretable predictive modeling: factor-level attribution, counterfactual analysis, and human-auditable internal representations for high-dimensional behavior-log data. The application domain is burnout risk from workplace signals such as meetings, email, and collaboration traces. The models are built so that a person can audit the predictions and reason about counterfactuals directly, rather than the system automating a decision on their behalf. Across Fall 2025 and Fall 2026 I have been Teaching Assistant to Prof. Yann LeCun and Prof. Alfredo Canziani for Deep Learning (DS-GA 1008). This fall I am also a Course Assistant for Applied Research in Discrete Optimization (IE-GY 9113A) at NYU Tandon under Prof. Kimberly Villalobos Carballo, and an instructor at Lens Academy.
Before NYU I completed a B.Tech in Computer Science (AI and ML) at SRM under Prof. Alice Nithya, with visiting research at UCL (Prof. Philip Treleaven), the University of Melbourne DISNET Lab (Prof. Adel Toosi), and project programs at Oxford (parameter-efficient fine-tuning under Pramit Saha) and London Business School. I co-founded HydroMind, which built and sold ModVPN and was acquired by 6th Element Inc.
Research
My research focuses on world-model interpretability, mechanistic interpretability, and AI safety. I am currently building World-Model-Lens, an open-source toolkit for world-model observability, and my COPHEE Lab thesis develops interpretable predictive models for high-dimensional behavior-log data.
- World-Model Interpretability. Capability-typed adapters, causal tracing across recurrent state-space models, token transformers, and JEPAs.
- Mechanistic Interpretability. Activation caching, circuit extraction, probing, and internal-representation analysis for learned systems.
- Interpretable Predictive Modeling. Factor-level attribution, counterfactual analysis, and human-auditable representations over high-dimensional behavior-log data.
- AI Safety. Open safety-auditing tools, OOD detection modules, hallucination diagnostics.
Stuff I’ve made
- World-Model-Lens. A mechanistic interpretability toolkit for world models: causal tracing, information-theory analysis, dynamical-systems tooling, multi-agent theory-of-mind probing, and SHAP / LIME modules.
- One Lens, Many Worlds: A Capability-Typed Interface for World-Model Interpretability. Capability-typed adapters that let one analysis run across RSSMs, token transformers, and JEPAs (2026).
- APEX: Adaptive Per-token Expert Loss. A per-token loss for MoE routing that adapts expert allocation on the fly (2025).
- Never Lost in the Middle Again: Teaching LLMs to Care About the Center of Long Documents. Mitigating the “lost-in-the-middle” failure mode of long-context LLMs (2025).
- Neural Seq2Seq with Attention for Abstractive Summarization. IJMLC, 2024. 22 citations.
- End-to-end Neural Embedding Pipeline for Large-scale PDF Retrieval. Distributed FAISS with Sentence Transformers, 2024. 17 citations.
- Efficient CAPTCHA Image Recognition with CNN + LSTM. IJARISE, 2024. 22 citations.
- Multi-Lead ECG Arrhythmia Detection. IJFMR, 2024. 10 citations.
- Dynamic Adaptation of Genetic Algorithms and Deep Learning in NLP. 2024. 10 citations.
- Financial Advisory LLM for Financial Literacy in India. 2024. 5 citations.
Teaching
- Deep Learning (NYU Courant, DS-GA 1008), Fall 2025. Teaching Assistant to Prof. Yann LeCun and Prof. Alfredo Canziani. Course site.
- Deep Learning (NYU Courant, DS-GA 1008), Fall 2026. Graduate Student Assistant to Prof. Yann LeCun and Prof. Alfredo Canziani.
- Applied Research in Discrete Optimization (NYU Tandon, IE-GY 9113A), Fall 2026. Course Assistant with Prof. Kimberly Villalobos Carballo.
- Forecasting, Modeling, and Future AI (Lens Academy), Fall 2026. Instructor. Cohort of ~20.
Contact
Reach me at any of these. I’ll usually reply within a day or two.
- Work / Personal work.bhavithchandra@gmail.com
- NYU bc4066@nyu.edu
Elsewhere
- Google Scholar. 86 citations · h-index 5 · i10-index 5.
- GitHub