Reliable text and speech generation
Making generated text and speech faithful to their source across languages, including languages with fewer resources.
Research interests
Safety of multimodal models
How vision-language models hold up against prompt injection and jailbreak attacks.
Trustworthy ML for health and brain signals
Models for medical images and EEG that can be interpreted and evaluated honestly, with an emphasis on how they are tested.
How I work
- Split data the way the model will meet it in practice: by participant for EEG, by patient for chest radiographs.
- Check generated output against its source, using factual entailment for summaries and consistency with the image and clinical record for reports.
- Pair automatic metrics with expert human ratings when the task is subjective.
- Use interpretability tools such as Grad-CAM and occlusion analysis to see what a model actually relies on.
Conferences, talks and service
Agentic AI Confluence 2026: Rise of the Orchestration Economy
Presented the multilingual summarisation paper at the NASSCOM Agentic AI Confluence.
Reviewer, CE2CT-2025
Reviewed submissions for the First Conference on Advances in Computer Science, Electrical, Electronics and Communication Technologies at Graphic Era Hill University.
8th International Conference on Innovative Computing and Communication (ICICC 2025)
Presented the paper “Bridging the Gap: Converting Read Text to Conversational Dialogue”.
Open to PhD research positions and Master's programmes
If you supervise work on multilingual NLP, speech, multimodal safety or medical AI, or you have an opening, I would be glad to hear from you.