About Me

Studio Quality Generative Music, Truly Controlled by the Artist. 

We're building an AI Music platform that goes much beyond text-to-music, and allows musicians to control every aspect of their music-creation process, including their sound, tone and song structure. 

Worked on LLM/VLM reasoning, Tiny LMs (LMs of the order of ~1B params) and cognitively inspired AI.

Papers:

1. Dylan Hillier, Leon Guertler, Palaash Agrawal, Chen Ruirui, Bobby Cheng, Cheston Tan, “ Super Tiny Language Models”, 2024.

2. Palaash Agrawal, Shavak Vasania, Cheston Tan, “Can LLMs perform structured graph reasoning?”, 2024.

3. Palaash Agrawal, Heena Rathore, Cheston Tan, “Advancing Perception in Artificial Intelligence through Principles of Cognitive Science”, 2023.

4. Palaash Agrawal*, Haidi Azaman*, Cheston Tan, “STUPD: a synthetic dataset for spatial and temporal relation reasoning”, 2023.

Worked on developing Computer Vision based products to improve the quality of mobility and monitoring for the elderly.
I was responsible for publishing two flagship AI modules (Locomo Test and Balance evaluation). Apart from my work in AI, I got the chance to work on some other interesting things in the short 6 months I worked there. For example, I designed the early drafts of the website experience (including payment experience), and lead the US FDA registration process. 

Worked on comparing deep neural networks with the visual recognition mechanism of the human brain through Brainscore - an aggregate of various metrics and neuroscience benchmarks. Advisor: Post-Doctoral Researcher Dr. Kamila Jozwik, PI: Dr. Jim DiCarlo.  

Working under Dr. Heena Rathore on redesigning AI Learning systems through cognitive principles. The Goal -- building Actor-Critic based Reinforcement Learning models inspired by Dopamine prediction and activation theory, and develop AI modeling principles utilizing mathematical models of perception, language and memory. 

1. Developed a deployment ready Machine Learning system for scalable tumor gleason grading classifier for existing medical imaging machinery (Xray Machines, PET/CT Machines, MRI machines, etc), with over 97% overall accuracy and 85% sensitivity over indolent cancer cases.

2. Worked on Data Engineering for CT scans extracted from Segmentation masks of cancerous regions. Introduced methods to overcome class imbalance and achieve model stability in extremely sparse and sensitive data. 

Worked as the principal Contributor and first author to a Computer Vision Research Project on Group Activity Recognition, under Dr. Ehsan Adeli from Stanford University 

Our paper, "LoTA: Local Temporal Attention Mechanism for Group Activity Recognition" is currently under review at BMVC 2021.

We achieved State of the Art result on 2 benchmark datasets (Collective Activity Dataset, Volleyball Dataset), while achieving the lowest time/space computational complexity among all related works. 

Led a team working on the project of Agricultural Boundary Segmentation of satellite images, using Deep Learning, under Dr. M.B. Potdar (Project Director at BISAG, and a former ISRO scientist). We successfully built a pixel-level binary segmentation CNN-based UNet, 

1. pretrained on open-sourced agricultural land satellite images, and 

2. fine-tuned on a small agricultural dataset (curated and manually labeled by our team) from various locations in India over various seasons of the year and times of the day.
   
   

This project's goal was to help resolve boundary disputes in certain rural areas of the state of Gujarat through AI systems.