Job Experience
DeepNeuronic
Software Engineer / ML Researcher - 1 Year:
- Created a custom theft action dataset and benchmarked hundreds of Large Video Language Models (LVLMs) for theft detection in surveillance applications.
- Researched and implemented multi-camera tracking methods to improve robustness across multiple views..
- Built a large-scale age and gender dataset using public, manual annotated, and synthetic data generated via generative models; trained and deployed an age and gender classification model.
- Created and annotated a personal protective equipment (PPE) dataset and trained a computer vision detection model to ensure compliance in real-world scenarios.
- Optimized core code for efficiency, reduced dependencies, vulnerabilities and documented the codebase for maintainability and onboarding.
Projects and Publications
Survey of Demonstration Learning
Comprehensive survey of demonstration learning, covering problem formulation, learning methods, benchmarks, applications, and open research challenges.
Hierarchical Decision Transformer
Advanced state-of-the-art sequence modeling in offline reinforcement learning using a hierarchical architecture.
- High-level and low-level transformers for task decomposition.
- High-level model proposes milestone states for the low-level model to reach.
- Milestones guide the policy to achieve complex tasks without prior knowledge or user interaction.
Decision Mamba Architectures
Further advanced the state-of-the-art sequence modeling in offline reinforcement learning by employing the Mamba architecture.
- Decision Mamba and Hierarchical Decision Mamba improve upon transformer predecessors in performance, size, and inference speed.
- Decision Mamba offers the best performance, while only employing one model.
- The evolutionary parameter provides the guidance signal missing in transformers.
DEFENDER
Algorithm to increase the safety of RL policies.
- Significantly reduces crash rates while improving overall policy performance.
- Uses small sets of safe and unsafe demonstrations.
- Compares the current trajectory with each demonstration.
- Terminates unsafe trajectories and disincentivizes dangerous behavior.
- Significantly reduces crash rates while improving overall policy performance.
Multi-View Contrastive Learning from Demonstrations
Framework to learn task-relevant visual representations from demonstrations.
- Leverages demonstrations recorded from multiple viewpoints.
- Trains an image encoder to produce viewpoint-invariant features.
- Uses these features as environment states to train reinforcement learning policies.
- Policies perform across unseen viewpoints and can extend to variations in illumination or background.
Music to Dance as Language Translation
Transformer and Mamba based framework to translate music to dance.
- Developed two Architectures: Mamba and Transformer based.
- Conducted ablation study on the effects of different music features.
- Proposed an algorithm to translate to translate music features to dance poses.
- Applied to a UR3 robotic arm and generalized to unseen choreographies.
About Me
I am a Machine Learning researcher and engineer with a Ph.D. in Computer Science, specializing in demonstration and reinforcement learning. I have published work in top peer-reviewed venues. I have 1+ year experience of industry experience applying ML models in production environments.
I have experience developing and deploying computer vision and ML models in production, working with large-scale datasets, and evaluating state-of-the-art models across robotics and surveillance domains.
I am particularly interested in ML engineering applied research roles.
My Resume
