News

• March 2025

  Check out new my blogpost about Constitutional Classifiers!

Industry Experience

• Research Engineer; University of California, San Francisco - San Francisco, CA

  September 2023 – Present

  Developed a multimodal method to improve interpretability of black box models by generating and refining concept bottleneck models within a Bayesian framework for uncertainty quantification, outperforming baseline methods by 35%

  Created a method to use a Bayesian tree based model that merges Llama 2’s domain knowledge with empirical data, increasing interpretability by 40%, as confirmed by domain experts, while matching top predictive performance and quantifying risk

  Designed and deployed a scalable ETL pipeline to process health record data, enabling ML training and evaluation for 3K+ patient records and 30K+ patient visits

  Built, deployed, and evaluated a custom clinical risk prediction model adopted by 10+ clinics, achieving 12% higher accuracy than the general clinical model

  Mentored 5+ peers and PhD students through teaching sessions and code reviews

• Software Engineer; Edovo - Chicago, IL

  Jan 2020 – May 2021

  Architected, tested, and deployed an educational content platform using Elasticsearch to handle 700K+ requests per day

  Led 10+ requirement gathering sessions with Product owners to re-build a platform

  Created a data pipeline and job to merge 4B rows of user event data in PostgreSQL

• Lead Software Engineer; 8th Light - Chicago, IL

  Aug 2017 – March 2019

  Implemented and deployed a scalable load testing platform simulating 1000+ RPS

  Engineered API integrations to sync 1000+ interactions/ minute in different timezones

Papers

• Concept Bottleneck Models with LLM Priors

  Jean Feng, Avni Kothari, et al

  ICML under review, 2024

  This work eliminates the need for human-annotated concepts by proposing a novel method to learn concepts by wrapping LLMs within a Bayesian framework. This approach is highly generalizable across various data modalities and allows for rigorous uncertainty quantification despite LLMs being prone to error and hallucinations.

  Prediction without Preclusion: Recourse Verification with Reachable Sets

  Avni Kothari, et al

  ICLR – Top 5% among submissions , 2024

  Individuals can be assigned predictions that they cannot change through actions on their features. This paper investigates and formalizes scenarios of predictions without recourse. We argue the importance of these scenarios for both model development and recourse detection methods.

  Bayesian Priors From Large Language Models Make Clinical Prediction Models More Interpretable

  Avni Kothari, et al

  AMIA - American Medical Informatics Association, Podium Abstract, 2024

  Implementing a Predictive Model to Reduce Hospital Readmissions in a Safety Net Healthcare System

  Arturo Gasga, Avni Kothari, et al

  ML4H - Machine Learning for Health, Oral Spotlight , 2024

Research Software

• bc-llm

  June 2024

  Implemented and developed a multimodal method using Metropolis Gibbs sampling to identify interpretable features for complex models

  Benchmarked and implemented 5+ comparator methods against our method, achieving performance comparable to or exceeding black-box models

• reachml

  June 2022

  Constructed a Mixed Integer Program to handle 50+ feature constraints for counterfactual explanations and test for robustness

  Created a model-agnostic fairness and safety audit to identify scenarios of preclusion

  Developed an HPC-based experimental pipeline to audit 200K+ individuals and benchmark results against baseline methods

Skills

ML Engineering

• ML Pipelines and ETL
• Multimodal datasets
• ML Deployment
• ML Development & Evaluation

AI Safety Research Skills

• Interpretability
• Robustness
• Fairness and Bias
• Risk Quantification
• LLMs and Foundation Models
• LLM Evaluations

Tools and Frameworks

• Python (Hugging Face, Pytorch, Scikit-learn, Numpy, Pandas)
• AWS (EC2, S3, Terraform, Deployment Strategies)
• Elasticsearch
• DB (SQL, Postgres, DuckDB)
• Docker

Education

• University of California, San Diego

  Masters in Computer Science

  2021 – 2023

  Thesis: Foundations of Model Agnostic Recourse Verification

• University of Texas at Austin

  Bachelors in Mathematics and Economics

  Minor: Computer Science

  2011 – 2016

Poster Presentations

• NeurIPS Workshop Statistical Foundations of LLMs and Foundation Models

  December 2024

• ICML Workshop on Data-centric Machine Learning Research

  July 2023

• ICML Workshop on Spurious Correlations, Invariance and Stability

  July 2023

• ICML Workshop on Artificial Intelligence & Human Computer Interaction

  July 2023

Teaching Experience

• Fall 2022

  TA for DSC 291 - Interpretability and Explainability in Machine Learning

• May 2011 – August 2013

  Differential Calculus Tutor

Service

• Vision 1948 (May 2023 – Present)
• PenPal for the Incarcerated (September 2020 – Present)
• Warren Community Garden (August 2021 – Present)
• UCSF AI4All (July 2024 – July 2024)
• The Recyclery (August 2018 – May 2021)