Architected and deployed a multi-agent AI research pipeline from scratch using Deep Agents (YAML-config-agent SDK), implementing a deterministic state machine with middleware-driven routing, reducing end-to-end media monitoring execution from 45 minutes to under 20 minutes. Designed a YAML-first framework abstracting agent lifecycle management, tooling, and execution logic enabling launch of systems in under 10 lines of code. Built a relevance-scoring merge layer with a Qdrant-backed vector search, Serp API asynchronous tool wrappers, URL deduplication, data filtering, and LLM scoring (threshold > 60). Engineered a production conversational AI system powered by Qwen 3.5 (deployed on Modal) on AWS SQS FIFO queues with DLQ fallback and a Qdrant-backed knowledge graph with user info for personalized experiences; integrated LangSmith and AWS S3 session management to gain actionable insights into user behavior and chatbot failure points.

AI Engineer at BMO (Contract Via Paceport)

March 1, 2025 - January 1, 2026

Built multi-agent AI systems using AWS AgentCore and LangGraph, implementing planner-executor architectures, API-based tool integrations, and context-aware orchestration to automate test generation workflows. Built an AI-powered design-to-code platform using AWS Bedrock, LangChain, and LangGraph to generate production-ready React components, reducing frontend effort by 40%. Built RAG services using LlamaIndex and multiple LLM providers (OpenAI, Anthropic, Mistral), implementing retraining, prompt chaining, multi-step retrieval, and fallback mechanisms reducing retrieval latency from 12s to 5s. Implemented reliability controls for LLM systems including structured output validation, fallback model chains, retry policies, and response quality guardrails to improve production stability. Worked directly with business stakeholders to translate workflow requirements into deployable AI automation systems.

Collaborating Researcher at MILA – Quebec Artificial Intelligence Institute

June 1, 2024 - December 1, 2024

Fine-tuned LaBRa M, a transformer-based Brain Model, to learn EEG representations for downstream tasks. Integrated neural tokenization and transformer-based EEG embeddings into the Speech Brain framework to support EEG-driven language tasks. Implemented multiple transformer and deep-learning EEG models (EEGNet, EEG Conformer, ShallowNet) with systematic hyperparameter optimization using Orion. Ran large-scale experiments on EEG samples, achieving 33.7% accuracy, outperforming state-of-the-art models by 2% and establishing a new benchmark for imagined speech recognition.

Associate Consultant - II at EY GDS

January 1, 2021 - December 1, 2022

Built production-grade classification systems using Logistic Regression, Random Forest, XGBoost, LightGBM, Decision Trees, GLMs, and Autoencoders for fraud detection, churn, and claims analytics, achieving 10–35% accuracy lifts and measurable cost savings. Developed automated ML deployment pipelines covering ETL, feature engineering, training, validation, deployment, monitoring, and retraining using AWS SageMaker Pipelines and Lambda. Containerized ML workloads with Docker and orchestrated production deployments on Kubernetes (EKS) for robust, fault-tolerant model serving with low-latency inference.