AI researcher · ML robustness

Hello, I'm Harshitha Machiraju.

AI researcher with a PhD in Machine Learning. I love building ML systems that actually work in the messy real world, from computer vision to NLP and now LLM safety. My focus? Making AI both clever and reliable.

Harshitha Machiraju, AI researcher and machine learning engineer

Experience & Education

Roles and degrees on one timeline, most recent first.

Jul 2026 – Present

LASR Labs Fellow

London, UK

Red-teaming LLM-as-judge reward models through RL and steering. Supervised by Andrew Draganov.

Apr 2026 – Present

AI Safety Expert & Mentor

BlueDot Impact · Remote

Working on AI alignment, robustness, frontier AI risks, and LLM safety evaluations.

Jan 2025 – Present

Independent AI Engineer

Remote

Designed and trained scalable LLM-based systems for health diagnostics and personalized shopping startups.

Sep 2019 – Nov 2024

PhD in Machine Learning

EPFL, Switzerland

Adversarial robustness and distribution shift: designing vision models that hold up when conditions get messy. Advisors: Prof. Pascal Frossard & Prof. Michael Herzog.

Sep 2018 – Aug 2019

Research Assistant

IIT Hyderabad, India

Developed and deployed ML models for autonomous navigation, including the implementation of adversarial testing frameworks.

Jul 2014 – Aug 2018

B.Tech in Electrical Engineering

IIT Hyderabad, India

Where I found algorithms and machine learning. Graduated summa cum laude with a minor in Computer Science.

Full resume →

Papers

Towards Robust Vision Models PhD Thesis, EPFL 2024
Doctoral thesis on designing vision models resilient to adversarial attacks and real-world corruptions.

Thesis
EREN: Enhancing Deep Learning Robustness through Image Pre-processing Under Review, 2024
A differentiable pre-processing method to defend models against natural corruptions.

Corruption & Pre-processing
Frequency-based Vulnerability Analysis of Deep Learning Models Under Review, 2023
Shows how spectral biases create vulnerabilities to corruptions and adversarial noise.

Corruption & Pre-processing
CLAD: Contrastive Learning for Background Debiasing BMVC 2022
Contrastive learning method that removes background bias, achieving SOTA on debiasing benchmarks.

Bias & Fairness
Empirical Advocacy of Bio-Inspired Models for Robust Image Recognition CVPR NeuroVision Workshop 2022
Makes the case for bio-inspired approaches as promising alternatives for robustness.

Bio-Inspired Models
A Little Fog for a Large Turn WACV 2020
First GAN-based adversarial fog attacks for stress-testing vision models in navigation.

Adversarial Robustness
Harnessing Latent Layer Vulnerabilities in Adversarially Trained Models IJCAI 2019
Introduces latent-space perturbations as a novel adversarial training method.

Adversarial Robustness
An Evaluation Metric for Object Detection in Autonomous Navigation ICIP 2018 (Oral)
Proposes a metric to evaluate object detection under varying weather conditions.

Evaluation Methods

Full list on Google Scholar.

Projects

Research projects from my PhD and before: adversarial robustness, corruption robustness, and the ways vision models quietly rely on the wrong features. Each card opens a write-up.

MUFIA

A multiplicative frequency attack that finds the spectral blind spots deep models share, and shows why "robust" models still fail on everyday corruptions.

Frequency analysis2023

A Little Fog for a Large Turn

Generative fog as a natural adversary: physically plausible weather that steers self-driving models off course without a single unnatural pixel.

WACV 2020Autonomous driving

CLAD

Contrastive learning that pulls classifiers off the background and back onto the object. State of the art on background-bias benchmarks.

BMVC 2022Bias & fairness

Latent Adversarial Training

Adversarially trained networks are still fragile inside. Attacking and then hardening the latent layers buys 4–6% robust accuracy for free.

IJCAI 2019Adversarial robustness

EREN

A differentiable pre-processing front-end that restores corrupted inputs before they reach the classifier, with no retraining of the backbone required.

Corruption robustness2024

Adversarial Subspace Analysis

Low-dimensional subspaces carry most of the discriminative signal in large language models. Here is how we found them.

LLMsInterpretability

Fairness vs. Robustness

Does hardening a model against attacks also make it fairer? We test the assumption and find a three-way trade-off instead.

Bias & fairnessAnalysis