I'm a Research Scientist at Meta Superintelligence Labs Trust ( formerly GenAI Trust, formerly Responsible AI) interested in safety alignment and interpretably of LLMs and AI for social good ( healthcare/public policy/general knowledge ) and have had my research published in ACL, CoNLL, ECML, EMNLP, ICLR, ICML, IJCAI and NeurIPS amongst other venues.
My research interests broadly revolve around (i) safety alignment and evaluation/mitigation of PII, IP, Biometric, Hallucicnation and Memorization risks of pre and post trained LLMs and (ii) interpreting and steering Machine Learning model decisions on natural language and multi-modal data, and learning representations of entities for downstream tasks. My PhD defended in July 2022 with my advisor Joydeep Ghosh and committee memebers Alex Dimakis, Harris Vikalo, Atlas Wang & Byron Wallace is titled "In-process Diagnostic methods for Entity Representation Learning on Sequential data at Scale" and focuses on methods that allow neural networks to be more transparent, explainable, and diagnosable during the process of learning and inference as opposed to in a post-hoc analysis fashion. My research has dealt with hallucination reduction with contrastive activation steering (NeurIPS 25 Mech Interp Workshop), measuring AI slop in text (under submission), memorization safety evaluation of Llama 3 OSS models, clustering influence-based embeddings for improved error analysis (NeuIRPS 23), open source tooling for explaining LLMs (EMNLP 23), intermediate entity-based sparse interpretable representation learning (EMNLP 22), efficient entity based knowledge injection for VQA (WWW 222),interpretable biomedical text representations (ACL 21), dense entity retrieval using dual encoders (CoNLL 19), prototypical learning of time series data (ICML 19(short) IJCAI 19 (long), efficient entity-based knowledge injection for KBVQA amongst other things.
key words: safety alignment, contrastive steering, influence functions, label quality, data valuation, memorization in LLMs, interpretable entity representations, knowledge injection for multimodal VQA, dense retrieval, in-network prototype learning, dual encoders, feature importance methods, counterfactual explanations
recent news
- Our work Hallucination reduction with CASAL: Contrastive Activation Steering For Amortized Learning has been accepted at NeurIPS 2025 Mechanistic Interpretability workshop
- Our benchmark paper Genµ: The Generative Machine Unlearning Challenge was accepted at ICCV 2025
- Our work on Measuring AI “Slop” in Text is now available on arXiv along with an overview by Chantal Shaib here
- Our work on "Improving LLM First-Token Predictions in Multiple-Choice Question Answering via Response Prefilling" is now available on arXiv
- I'm a co-organizer for the 2nd Unlearning and Model Editing (U&ME’25) workshop at ICCV’25, Hawaii, Oct 2024. Call For Papers Deadline July 1, 2025!
- I was part of the great team that worked on Meta's recently launched open-source LLama 4 in May 2025
- I'll be hosting a Research Scientist intern for 2025 with a focus on exploring LLM safety alignment and understanding potentially using explainability methods. More info here
- I was the primary co-organizer for the Unlearning and Model Editing (U&ME’24) workshop at ECCV’24, Milano, Italy, Oct 2024 which was a great success and David Bau and Rita Cucchiara were wonderfull keynote speakers
- I was part of the great team that worked on Meta's recently launched open-source LLama 3 LLM in April 2024 and was a core contributor to the Llama 3.1 release in July 2024
- Our paper "Error Discovery by Clustering Influence Embeddings" was accepted at the 2023 NeuIRPS main conference. An earlier version of this paper was previously accepted as an Oral presentation at the ICLR 2023 Trustworthy ML workshop
- Our paper Using Captum to Explain Generative Language Models at the EMNLP 2023 - NLP OSS workshop!
- I was an invited speaker and panelist at Health Day @ The Web Conference 2023 where I presented a talk on Explainable Machine Learning with applications in Biomedicine and Healhtcare and discussed topics including generative AI and LLMs (large language models).
- Our paper "Intermediate Entity-based Sparse Interpretable Representation Learning" was accepted into the the BlackboxNLP workshop at EMNLP 2022. Here is the poster I presented and our github code
- As of Summer 2022, I'm a PhD graduate in Electrical & Computer Engineering (ECE) at UT Austin with my advisor Dr. Joydeep Ghosh. My PhD defense slides are here.
- March 2022, I was accepted as a Fulbright Specialist to work with the Center of Human Rights at UNICEN in Azul, Argentina on predicting gender bias in judicial proceedings using NLP. In addition to the researchers at UNICEN, I will be working with Dr. Maria De-Arteaga. As part of the project, we setup a workshop on AI and Public Policy/Law in Spanish. Our paper "Detección de sesgos en razón del género en decisiones judiciales utilizando PLN" was accepted for publication based on work during that time.
- Feb 2022, "Improving and Diagnosing Knowledge-Based Visual Question Answering via Entity Enhanced Knowledge Injection" has been accepted at ACM WWW 2022 in the Workshop on Multimodal Understanding for the Web and Social Media. Here are the slides for an invited talk on our work.
- Dec 2021, preprint on "Improving and Diagnosing Knowledge-Based Visual Question Answering via Entity Enhanced Knowledge Injection" is available. Paper || Code
- Summer 2021, my PhD Proposal on "In-process Diagnostic methods for Entity Representation Learning on Sequential Data" has been accepted for candidacy.
- Our paper "Biomedical Interpretable Entity Representations" has been accepted at ACL-ICJNLP 2021. Paper || Code || Slides
- For the Summer 2020, I have been selected to be an IBM Research PhD fellow for Social Good and will work on a Drug Repurposing for Cancer project using NLP.
- I recently gave a high level talk to Cognitive Scale on Explainable AI for NLP (slides link).
- In Summer 2019, I interned at Google Cloud AI in Seattle with Besim Avci and Frederick Liu working on explaining seq2seq models using feature attribution methods on Transformer and LSTM based architectures with attention for machine translation.
- In Spring 2019: I was a TA for Responsible AI Graduate Seminar ( lots of papers and presentations on Explainability, Fairness, etc!)
- In Summer 2018, I interned at Google Research in Mountain View with Jason Baldridge and Daniel Gillick working on entity linking.
- In Fall 2017: I started at UT and was a TA for an Advanced Data Mining Masters course.
- In Summer 2016, I was selected as an Eric & Wendy Schmidt Data Science for Social Good 2016 fellow at the Unviersity of Chicago and worked with SEDESOL of Mexico on improving their distribution of social services.
- I also do alot of Data Visualizations. Here is a lecture I have given on modern data visualization and particulary the d3 javascript library for masters level CS/ML students.
- I obtained a Masters of Data Science at the UPC in Barcelona in July 2015 (masters thesis on automated construction of political networks). I have a bachelors in Computer Science, Political Science and Hispanic Studies with a Business minor from UT Austin.
A selection of publications and projects, academic, professional and personal.
-- Paper || Project/Code
|| Llama 3 405B online
-- Paper || Code || Slides
-- Paper || Code
-- Paper || Code || Poster
-- Paper || Code
-- Paper || Code || Slides
-- Paper || Code
-- Paper || Code
IJCAI 19 Knowledge and Health Discovery Long paper
-- Paper || Code
Identify Blue Counties in America that are surrounded by Red ones, and Predict if a county is a blue island based on just socio-economic and public health data.
List of Spotify's Clarify Data Stories series articles written by Rob Mitchum for which I contributed data mining and data visualization, 2016.
'
How Music Crosses American Borders
Debt Anxiety in Today’s Pop Music
Visualizations related to my masters thesis project in Barcelona about Texas Politics. Click here for Thesis Presentation Slides
