CV
You can also download the PDF version above.
Contact Information
| Name | Harsh Kumar Agarwal |
| Professional Title | ML Researcher |
| harsh19423@iiitd.ac.in |
Professional Summary
ML researcher working on generative models, multimodal learning, and embodied AI. Currently a Research Intern at Five.ai in Cambridge, UK, following a Master’s in Artificial Intelligence at ENSIMAG/UFR IM²AG and thesis research at INRIA Grenoble.
Experience
-
2026 - present Cambridge, United Kingdom
Research Intern
Five.ai
- Research internship following the completion of my Master’s thesis at INRIA Grenoble.
-
2026 - 2026 Grenoble, France
ML Research Intern — RobotLearn Team
INRIA Grenoble
Master’s Thesis. Adviser: Dr. Xavier Alameda-Pineda.
- Flow-Matching Models for Multimodal Action Generation: Developed EMODY Flow, a flow-matching framework integrating visual, auditory, and linguistic inputs to produce emotion-conditioned full-body motion; accepted at the JAV-CG workshop, ACM Multimedia 2026.
- Implemented flow-matching decoders with cross-attention mechanisms to leverage Qwen-3 Omni audio responses for embodied robotic decision making and control.
- Trained on the BEAT2 dataset (SMPL-X body and FLAME facial representations).
-
2023 - 2025 Bangalore, India
Software/Firmware Developer
IBM
- Redundant BMC Architecture: Designed and implemented redundant Baseboard Management Controller (BMC) for POWER server systems. Expertise in low-level system robustness and fault tolerance.
- PEL Parser Porting: Ported Platform Event Log (PEL) parser from C++ to Python, enabling cross-system deployment and improving code maintainability.
-
2022 - 2023 New Delhi, India
Undergraduate Researcher
VISION Lab, IIIT Delhi
Adviser: Dr. Saket Anand.
- AI for Agriculture — Multi-Spectral Satellite Image Analysis: Analyzed crop parameters using multi-spectral satellite imagery via Google Earth Engine. Applied semi-supervised learning techniques; published dataset paper at WACV 2024.
- Hierarchical Semi-Supervised Image Classification: Extended HierMatch architecture to handle hierarchical label structures with unlabeled data at different hierarchy levels.
-
2022 - 2022 Remote
ML Research Intern
TCS Research
- Self-Supervised Learning for Mitotic Cell Recognition: Predicted cancer stages via cell division counting. Applied DINO to learn robust single-cell image features without extensive manual annotation.
-
2021 - 2021 Remote
ML Research Intern
STAQU Technologies
- Transformer-Based Action Recognition: Fine-tuned TimeSformer for video clip classification with real-time processing on streaming data.
- Novel Object Detection from CCTV: Developed change-detection system using image differencing, deployed on streaming video.
- Car Color Classification: Built k-NN classifier for 12-color car classification achieving ~80% accuracy.
Education
-
2025 - 2026 Grenoble, France
-
2019 - 2023 New Delhi, India
Bachelor of Technology
Indraprastha Institute of Information Technology Delhi
Computer Science, Specialization: BioScience
- GPA: 8.83 / 10
Publications
-
2026 EMODY Flow: Emotion-Aware Audio-Driven Full-Body Motion Generation
1st International Workshop on Joint Audio-Video Comprehension and Generation (JAV-CG), ACM Multimedia 2026
Accepted paper. A lightweight flow-matching framework that restores emotion sensitivity in audio-driven full-body motion generation, setting a new state of the art on BEAT2 and transferring zero-shot to facial animation.
-
2024 SICKLE: A Multi-Sensor Satellite Imagery Dataset Annotated With Multiple Key Cropping Parameters
IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) 2024
Accepted for oral presentation. Multi-spectral satellite dataset for key cropping parameters at 3m, 10m, and 30m resolutions.
-
2025 Cross-Cultural Recipe Transformation Via Neural Network and Encoder-Based Models
IEEE International Conference on Data Mining Workshops (ICDMW) 2025
Skills
Languages
Projects
-
Cuisine Transformation via Semantic Recipe Embeddings
Developed a neural network classifier for cuisine type identification from recipes. Created an unsupervised ingredient substitution method via semantic embeddings, achieving 90% accuracy in cuisine preservation after transformation. Published at ICDMW 2025.
-
Bird Species Classification from Audio
End-to-end audio classification pipeline using Mel-Spectrograms and deep learning. Achieved 70% accuracy on bird species identification. Code available on GitHub.