cv

PhD in Computer Science

Imperial College London

Expected completion 2027

• Conducting research on AI safety, adversarial robustness and model evaluation for machine learning systems under the supervision of Prof. Alessio Lomuscio, including conversational LLM agents, RAG systems, and decision-tree ensembles.
• Supported by the UKRI Centre for Doctoral Training in Safe and Trusted Artificial Intelligence.

MEng in Aeronautical Engineering

Imperial College London

• First Class Honours (78%), Dean’s List

Research Scientist (Part-time)

Safe Intelligence Ltd.

• Core contributor at an Imperial College London spinout for the product design and R&D of a SaaS product for testing and validation of AI agents.
• Designed iterative, optimisation-based methods for multi-turn jailbreaking of AI agents, achieving an attack success rate of up to 50% against Sonnet 5, GPT-5.6, and Gemini 3.6-powered agents evaluated on HarmBench behaviours, exposing alignment failures that emerge over extended interactions.
• Developed optimisation-based black-box methods that generate intent-preserving adversarial queries to stress-test RAG agents, eliciting evidence-contradicting intrinsic hallucinations in up to 54% of responses from GPT-5 and Gemini 3 models.
• Led the synthetic test-data generation workstream, designing methods to achieve broad domain coverage for systematic agent evaluation.
• Proposed and implemented a robust gradient boosting algorithm for tabular data models, improving robust accuracy by 44% over the state of the art and 113% over conventional XGBoost.

Product Engineer

Pangaea Data Ltd.

• Directed a four-person cross-functional NLP and product team to identify patients with rare diseases from electronic health records, collaborating with clinicians and stakeholders to refine research objectives and validate product design.
• Designed a clinical NLP pipeline that used BERT-based models and regex keyword detection to mine unstructured text, extract ontology concepts, and identify patients through a decision-tree model, distributed on Microsoft Azure Marketplace.
• Led a proof of concept for a major pharmaceutical company, applying the pipeline to data from approximately 65,000 patients to identify cohorts with cachexia, COPD, and lupus nephritis; the approach identified 2-5x as many patients compared to a diagnostic code baseline.
• Authored technical white papers, research grant submissions and project proposals leading to 2 POCs.

Intent-Preserving Adversarial User Simulation for Multi-Turn Agent Evaluation

UserSim Workshop at NeurIPS, 2026 (to appear)

Atri Vivek Sharma, Brian Formento, Alessio Lomuscio

Eliciting Intrinsic Hallucinations in LLMs via Semantically Equivalent Adversarial Attacks

Conference on Language Modeling (COLM), 2026

Atri Vivek Sharma, Brian Formento, Alessio Lomuscio

Learning Robust XGBoost Ensembles for Regression Tasks

Uncertainty in Artificial Intelligence (UAI), 2025

Atri Vivek Sharma, Panagiotis Kouvaros, Alessio Lomuscio