Armin – AI agent development, Python, LLM

Leading the AI Research & Development division focused on advancing automation and intelligence in professional translation workflows. The role involves designing, prototyping, and productionizing state-of-the-art AI systems for translation quality estimation, adaptive machine translation, and multilingual content optimization. The work bridges academic research and industrial deployment, implementing the latest findings from NLP and machine translation literature into scalable, production-grade solutions.

• AI Research & Implementation: Conducted applied research on emerging NLP and MTQE methods, transforming peer-reviewed techniques into deployable production systems tailored to translation workflows. • Model Development & Evaluation: Designed, fine-tuned, and evaluated transformer-based models (e.g., COMET, mBERT, GPT, custom encoder-decoder architectures) to enhance translation accuracy, fluency, and context preservation across multiple languages. • Pipeline Automation: Architected end-to-end automated pipelines streamlining translation, evaluation, and delivery, integrating translation memory, quality estimation, and linguistic review in a continuous AI-assisted loop. • Quality Estimation (MTQE): Led the design of an AI-driven translation quality estimation system combining reference-free and reference-based evaluation methods for consistent and explainable scoring. • R&D Leadership: Oversaw experimental design, benchmarking, and internal research documentation; provided technical direction for prototype validation and production integration. • LLM Integration & Contextual Translation: Developed and evaluated LLM-based translation enhancement and correction modules using OpenAI and HuggingFace APIs to improve domain-specific accuracy and tone adaptation. • Data Engineering & Retrieval: Implemented context-retrieval mechanisms using vector databases (Pinecone, ChromaDB, OpenSearch) to enhance translation consistency and maintain semantic coherence across large corpora. • Evaluation Frameworks: Integrated LangChain, LangSmith, and LangGraph to orchestrate multi-agent pipelines, ensure traceability, and automate evaluation of model outputs across production datasets. • Scalable API Infrastructure: Developed FastAPI-based microservices and REST endpoints supporting concurrent evaluation and model inference, deployed on Azure AI infrastructure with optimized resource management.

Led an MLOps maturity assessment for a large financial institution to modernize and standardize their model lifecycle using Vertex AI. The project focused on improving governance, automation, and monitoring for enterprise-scale AI adoption.

• Current State Analysis: Reviewed pipelines, deployment workflows, and model promotion practices to assess MLOps maturity. • Stakeholder Engagement: Conducted interviews with data scientists, architects, and AI governance leads to align technical architecture with compliance needs. • Architecture Design: Co-designed a scalable, secure Vertex AI–centered platform for training, validation, deployment, and monitoring of ML/LLM models. • Governance & Compliance: Helped define a model promotion strategy, including role-based access control, audit logs, and approval workflows. • Observability: Proposed and documented use of Prometheus + Grafana, Cloud Monitoring, and metadata logging for real-time tracking and drift detection. • Automation & CI/CD: Recommended and partially implemented automated retraining pipelines, model versioning, and CI/CD integrations using Terraform and GitHub Actions. • Documentation: Co-authored MLOps assessment reports, a harmonized roadmap, and playbooks for model lifecycle operations in highly regulated environments.

Contributed to the design and implementation of a future-ready MLOps architecture on AWS to automate and optimize machine learning operations for a client. The project involved assessing current ML pipelines and proposing a scalable, automated, and secure architecture for data ingestion, model retraining, deployment, and monitoring. As part of the initiative, an MVP was developed using 2–3 of the client’s existing models to validate the proposed architecture and demonstrate scalability. Key focus areas included addressing data pipeline inefficiencies, enhancing model management, integrating CI/CD for ML workflows, and adhering to strong security protocols.

• Current State Assessment: Reviewed existing data pipelines, model training flows, and deployment setups to identify scalability and automation gaps. • Data Pipeline Optimization: Designed and built data ingestion pipelines supporting real-time and batch processing, with scalable storage and data quality monitoring for selected models. • Model Pipeline Development: Implemented retraining workflows, hyperparameter tuning, and model versioning using SageMaker Training Jobs and Model Registry. • Infrastructure Modernization: Architected compute infrastructure for scalable model training and deployment using AWS services such as SageMaker, RDS, and IAM. • Automation & CI/CD Implementation: Established CI/CD pipelines for seamless deployment of data pipelines and ML models using Terraform for infrastructure as code. • Model Deployment & Monitoring: Deployed models with real-time and batch inference, validated scalability via autoscaling endpoints, and implemented automated testing for pipeline consistency. • Security Enhancements: Enforced security measures including RBAC, encryption for data storage, and IAM-based access controls to ensure compliance. • Future State Architecture Design: Defined a scalable, automated ML pipeline architecture supporting continuous integration, retraining, and governance for the MVP. • Tooling & Automation Gaps Resolution: Addressed gaps in existing CI/CD tooling to enable end-to-end automation for ML operations and data pipeline management. • Scalability Validation: Validated architecture scalability through extensive testing of real-time inference pipelines and model performance under varied loads.

Project 1 – Feature-Gen (Academic): Created a hierarchical two-level genetic-algorithm framework that automates feature selection and transformation; published findings at a journal and conference and released the open-source feature-gen library for ML optimization.

Project 2 – Structured-Data Reasoning Agent (Mitacs × industry partner): Built an LLM-driven system that converts natural-language questions about tabular data into SQL or ML analyses, automatically inferring inter-column relationships with different visual dashboards.

Project 3 – ImmiAI Chatbot (Mitacs × industry partner): Delivered a RAG-powered assistant that scrapes authoritative immigration sources, indexes them in a vector store, and returns grounded answers to users planning to move to Canada.

• Algorithm R&D: Designed and coded a two-level NSGA-II genetic algorithm with bootstrapping, transformation pools, and ensemble scoring; packaged it as feature-gen and authored the publication. • LLM Architecture & Integration: Orchestrated GPT-4–based pipelines with LangChain to parse user intent, map schema, and route queries to SQL engines, statistical tests, or XGBoost models. • LangGraph Usage: Used LangGraph to create node layers sharing state such as dataset information and metadata. • Prompt Engineering: Developed few-shot, ReAct, and chain-of-thought prompts; versioned them in Git and ran A/B experiments to optimize accuracy, latency, and token cost. • Embedding & Tokenization Strategy: Selected and configured OpenAI and HuggingFace embeddings; handled context limits, token chunking, and compression to optimize semantic search accuracy and cost efficiency. • Fine-Tuning & Adaptation: Conducted LoRA-style experiments adapting pretrained LLMs on domain-specific immigration and tabular query datasets while balancing compute overhead and response quality. • Retrieval-Augmented Generation: Implemented a Pinecone/OpenSearch vector layer embedding scraped IRCC content and chat history with search-time reranking and citation injection. • Vector Database Evaluation: Benchmarked OpenSearch, Pinecone, and ChromaDB for RAG pipelines based on latency, scale, multi-tenant support, and embedding compatibility. • LLM Cost & Latency Optimization: Tuned temperature, top-p, and max tokens across prompt versions. • Evaluation & Testing: Built automated suites measuring perplexity, ROUGE-L, BLEU, custom relevance scores, and schema consistency; logged metrics to MLflow. • Monitoring & Observability: Instrumented FastAPI services with Prometheus exporters and Grafana dashboards for uptime, token usage, latency, and drift alerts. • Security & Compliance: Implemented JWT authentication, IAM-scoped AWS resources, and encryption at rest and in transit to meet academic and partner data-protection requirements. • API Design for LLM Interfaces: Exposed LLM systems via FastAPI and WebSocket APIs for real-time interaction through a Next.js frontend, supporting structured queries, file uploads, and streaming responses.

Led the development of a cloud-native SaaS platform that automates large-scale SEO audits across 50,000+ links per domain. The system extracts actionable BI insights through advanced web-crawling and NLP pipelines. Architected as a distributed microservices network, it uses LLMs and prompt engineering to generate human-readable summaries and explanations of SEO metrics.

• Advanced Data Analysis & Extraction: Employed ML and NLP to distill BI from large datasets using state-of-the-art analytical techniques. • Real-time Data Communication: Implemented WebSocket protocols to enable instantaneous data exchange among microservices, improving responsiveness and UX. • Innovative Application Development: Utilized LangChain to develop LLM-powered applications, leveraging AI technologies. • Performance Optimization: Deployed Redis for in-memory caching, accelerating data retrieval and processing speeds. • Frontend Refinement: Overhauled frontend architecture to optimize performance and mitigate memory leaks, improving UI responsiveness and user satisfaction. • Microservices Architecture: Developed and integrated 12+ microservices ensuring interoperability and resilience. • Parallel Programming Optimization: Optimized performance using parallel programming techniques. • Database Management: Used RDS and Firebase real-time DB, tailoring storage solutions to business requirements. • Security Enhancement: Implemented authentication mechanisms across microservices to strengthen security. • Legacy System Maintenance: Managed and updated a legacy Node.js project, ensuring continuity and integration with new technologies. • Scalability & Efficiency: Leveraged Kubernetes for efficient deployment of multiple processing pods. • Infrastructure Automation: Used Terraform for infrastructure as code to enable scalable cloud resource management. • API Design: Crafted RESTful APIs using OpenAPI specifications for standardized data exchange. • Cloud Solutions: Utilized AWS, Firebase, and GCP services to optimize costs while meeting operational requirements. • CI/CD with GitHub: Used GitHub CI/CD pipelines to automate deployments and streamline workflows.