Machine Learning Engineer

BHAVIK ARDESHNA

+1-438-***-**** Mail: **************@*****.*** LinkedIn: Bhavik Ardeshna Github: Bhavik-Ardeshna EDUCATION

MILA - Quebec AI Institute, Montreal (Affiliated with Universit e de Montr eal) Sept 2023 - Dec 2024 Master’s in Computer Science with Specialization in Machine Learning Relevant Coursework: Representation Learning, Machine Learning, Data Science. Dharmsinh Desai University, India May 2019 - June 2023 Bachelor’s in Computer Science with Minor in Natural Language Processing (NLP) Relevant Coursework: Distributed System, Artificial Intelligence, System Design. SKILLS

Languages C/C++, Python.

Libraries/Frameworks: PyTorch, TensorFlow, ONNX, CuPy, FastAPI, Git, Docker, Kubernetes, OpenCV, Unstructured. Technologies: GCP, AWS (EC2, SageMaker, etc), Azure, MLFlow, Lambda Labs, HuggingFace. LLM-Frameworks: vLLM, Haystack, LangChain, Langgraph, Ollama, Milvus, LanceDB, Pinecone. EXPERIENCE

Senior Machine Learning Engineer - GenAI Dec 2024 - Present Binoloop Toronto, Ontario

• Leading the Atlas Agentic AI platform to streamline government procurement processes and evaluate large tender documents.

• Implementing a multi-agent framework utilizing LangGraph and LangChain. Conducted applied research with Binoloop to enhance LLM reasoning capabilities.

Machine Learning Engineer May 2024 - Dec 2024

Humanware Montreal, Quebec

• Working on developing diffusion-based models for text restoration, improving GPU latency for faster inference, and providing API support for other Humanware product lines.

• Optimized Swin-Transformer model inference on NPU, improving performance by 3% for Explore 2.0 Magnifier in super- resolution and image restoration of low-resolution, high-zoom images. Machine Learning Engineer, Intern Jan 2024 - May 2024 Sycodal Montreal, Quebec

• Enhanced object detection and custom segmentation models by 9% for Sycodal Robotics, reducing false pick-and-drop operations by 11.3% in custom automation tasks.

• Streamlined computer vision model training, registry, and deployment using MLflow, with inference on Azure edge-cloud and also perfrom A/B testing. Worked with robotic path estimation algorithms to automate tasks, reducing human intervention. Machine Learning Engineer Dec 2021 - Jul 2023

SoftamaxAI Ahmedabad, Gujarat

• Implemented video understanding tasks with the SlowFast model (by Meta) for action understanding, boosting performance by 7.8%. Designed a real-time camera streaming system with inference on NVIDIA Jetson via AWS.

• Fine-tuned LLaMA for text generation tasks, enhancing LLM performance and scaling on AWS EC2 for improved inference speed and workload scalability.

Machine Learning Engineer, Intern Dec 2020 - Apr 2021 Heliconia Solutions Ahmedabad, Gujarat

• Collaborated on a AI-powered customer service platform, leveraging transformer-based models to automate QA systems using information-retrieval. Integrated multilingual support into the SaaS platform, enhancing user experience. PUBLICATIONS

GujaratiQASuite: Novel Resources for Gujarati Question-Answering System. Developed GujQA, the baseline QA benchmark for Gujarati, achieving a 64.99 F1 score and 47.65 EM. Leveraged zero-shot cross-lingual transfer with GujAdapter, outperforming English-based models using Hindi as the base language. (Under Review) Cascading Adaptors to Leverage English Data to Improve Performance of Question-Answering for Low-Resource Languages. Enhanced low-resource language performance in question answering by +2% using cascading adapters with trans- former models. We trained 4 variants of adapter combinations for seven lanagues. Paper & (42.5K Downloads in HuggingFace) PROJECTS

Multimodal-VideoRAG Github

• Developed a multimodal information retrieval and question-answering system on videos using Video Retrieval-Augmented Generation. Leveraged Large Vision-Language Models (VLMs) and BridgeTower embeddings for video pre-processing, embedding generation, and multimodal vector database queries with LanceDB. Llama-CuPy Github

• Integrated CuPy with LLaMA for GPU-accelerated matrix operations, reducing inference time and improving efficiency over PyTorch implementations, achieving 2x performance improvement on large models with NVIDIA RTX 3090. RecLLM: Personalized Complementary Product Recommendation using LLM Github

• Developed an zero-shot complementary product recommendation system using LLM, transformer-based encoder for product encoding and LLM queries to generate and filter complementary product candidates based on customer shopping behaviors. Multi-Task Adapter based Question-Answering for-Low-Resource-Languages Github

• Demonstrated that at by using the transformer model with multi-task adapters, the performance of Question-Answering downstream tasks for low-resource languages, such as Hindi, Arabic, and Vietnamese, can be improved by leveraging high- resource language understanding.

CONTRIBUTIONS & ACHIEVEMENTS

• Kaggle 3x-Expert Contributed to a multilingual Visual Question Answering (VQA) task and achieved a top-10% finish in the American Express Default Prediction Hackathon. Kaggle

• Hugging Face: Developed and published 13 multilingual transformer models (mBERT, XLM-R) for seven languages, improving QA system performance in low-resource settings. (42.5K Downloads)

• Articles Basic intuition of Conversational Question Answering Systems (CQA) and Semantic Alignment of Linguistic and Visual Understanding using Multi-modal Transformer, was featured in Google at the Becoming Human: Artificial Intelligence Magazine.

VOLUNTEER EXPERIENCE

Team Coordinator Aug 2019 - Jan 2023

Anurakti Foundation Nadiad, Gujarat

• Led volunteer initiatives across 8 city camps, coordinating 50+ volunteers to teach 300+ students while organizing sports, cultural, and educational events, fostering holistic learning experiences for children.

Contact this candidate