Machine Learning Engineer building multilingual NLP systems, evaluation frameworks, and efficient AI architectures.

📍 Toronto, Ontario, Canada

LinkedIn • Hugging Face • The Meta Gradient

I'm a Machine Learning Engineer focused on multilingual NLP, evaluation systems, dataset engineering, and efficient machine learning architectures. My work spans open-source research, data pipelines, ML engineering, and technical writing, with an emphasis on building practical, reproducible systems grounded in real-world data.

I currently serve as Cross-Team Coordination Lead for the Multicultural Riddles Benchmark within the Cohere Labs Open Science Community, where I help coordinate multilingual dataset development, evaluation workflows, and benchmark infrastructure across a global contributor community.

Alongside open-source work, I independently research efficient sequence models, representation learning, reasoning systems, and optimization dynamics while publishing implementation-focused articles through The Meta Gradient.

• 🌍 Cross-Team Coordination Lead for the Multicultural Riddles Benchmark (Cohere Labs Open Science Community)
• 🤖 Building multilingual NLP datasets and evaluation pipelines
• 🧠 Researching State Space Models (SSMs), Mamba, and efficient neural architectures
• 📊 Designing scalable datasets, schemas, and ML data pipelines
• ✍️ Writing implementation-focused deep learning articles on The Meta Gradient

Cross-team coordination, multilingual data pipelines, and evaluation workflows for an open-source cultural reasoning benchmark spanning 49 languages, ~49,000 riddles, and 70+ contributors.

• Appointed Cross-Team Coordination Lead, coordinating across data creation, evaluation, and analysis teams.
• Developed internal tooling that converts validator output into actionable guidance for 70+ contributors, improving annotation quality and review efficiency.
• Designed and shipped the pilot dataset (v1) covering 40 communities and 40,000 riddles using a canonical schema for downstream benchmarking.
• Help standardize multilingual datasets for structured releases on GitHub and Hugging Face.
• Collaborate on benchmark design across API-based and open-weight language models.
• Support preparation for an upcoming open-science research publication.

Tech

Python • Pandas • Hugging Face Datasets • Git • GitHub • Schema Design • ISO 639-2/3 • Multilingual NLP

Repository (Private)

Fully offline multilingual document question-answering system for newcomers and migrants.

• Pipeline Lead on a 7-person team.
• Built an end-to-end RAG pipeline:
  • PDF extraction
  • BGE-M3 embeddings
  • ChromaDB retrieval
  • Tiny Aya generation
  • mDeBERTa hallucination verification
• Executed over 9,000 automated evaluations across Chinese, Hindi, and Polish.
• Identified cross-lingual embedding quality—not LLM capability—as the primary retrieval bottleneck.

Tech

Python • PyTorch • BGE-M3 • ChromaDB • Tiny Aya • mDeBERTa • Gradio

Repository

https://github.com/docunative-AI/docunative

Adaption Labs Uncharted Data Challenge

Created a high-quality Tamil agricultural instruction dataset from public-sector resources.

• Built the entire dataset independently from Kisan Call Centre, TNAU Extension Guides, and ICAR contingency plans.
• Curated 187 structured records covering 48 crops across 20 agricultural categories.
• Completed 10 iterative submissions over six weeks.
• Awarded Grade A (9.4/10) with honorary recognition from Sara Hooker.

Tech

Python • Pandas • Dataset Engineering • Metadata Design

Dataset

https://huggingface.co/datasets/vinod-anbalagan/tamil-agri-advisory-qa

Repository

https://github.com/VinodAnbalagan/tamil-agri-dataset-

Implementing neural architectures from first principles to better understand the evolution of sequence modeling.

Current implementations include:

• Vanilla RNN
• LSTM
• GRU
• Transformer
• State Space Models (Mamba)

Research focuses on:

• Gradient flow
• Memory mechanisms
• Computational efficiency
• Scaling behavior
• Architectural trade-offs

Tech

PyTorch • NumPy

Studying optimization dynamics and neural network convergence through visualization.

• Compare optimization behavior across architectures.
• Explore why certain models converge more reliably.
• Visualize training dynamics and loss surfaces.

Tech

PyTorch • Matplotlib

• Cross-Team Coordination Lead
• Multilingual benchmark development
• Evaluation pipelines
• Dataset engineering

Publishing datasets and machine learning resources.

Technical articles focused on deep learning, efficient architectures, and implementation from first principles.

• Python
• SQL
• Bash

• PyTorch
• Scikit-learn
• XGBoost
• Model Evaluation
• Feature Engineering

• Transformers
• Retrieval-Augmented Generation (RAG)
• Sentence Embeddings
• Hugging Face
• ChromaDB
• Multilingual NLP

• Pandas
• NumPy
• Dataset Curation
• Data Cleaning
• Schema Design
• Evaluation Pipelines

• FastAPI
• Docker
• AWS
• Git
• GitHub
• Linux

• Matplotlib

M.A.Sc. Electrical Engineering
University of Windsor

B.E. Electronics & Communication Engineering
Anna University

• University of Toronto Machine Learning & Data Science Certificate
• Stanford Machine Learning Specialization
• Google Advanced Data Analytics

I write implementation-focused articles that break down modern machine learning research into practical, reproducible code.

Topics include:

• Efficient neural architectures
• State Space Models
• Deep learning fundamentals
• Optimization
• Representation learning
• Machine learning engineering

The Meta Gradient

https://substack.com/@vinodanbalagan

• Multilingual NLP
• Efficient AI
• Representation Learning
• State Space Models
• Evaluation Frameworks
• Dataset Engineering
• Open-source ML
• AI for Scientific Research

"Learning by building, understanding by implementing, and improving through open collaboration."