Hello, I'm
Marium Ashraf
|
Hello, I'm
|
Get To Know More
3+ years
AI Systems, LLM Engineering & Full-Stack Development
B.Eng. Computer Engineering (Major in Software Engineering)
Toronto Metropolitan University
Graduating April 2027
I'm an Operations & AI Engineer building production-grade systems that combine LLM intelligence with distributed architecture. At Lyrata, I design AI-assisted tooling that reduces operational latency by 90% and cuts onboarding time in half. My work spans microservice LLM evaluation pipelines, vector database systems, and full-stack applications — all built with observability, fault tolerance, and scalability in mind.
From GPT-4o-powered RAG systems and MLflow experiment tracking to containerized microservices and real-time analytics dashboards, I build systems that teams can trust at scale. I've improved model quality by 35%, reduced debugging cycles by 60%, and accelerated feature delivery by 30% through architectural refactors and standardized workflows.
Key
SOP Lookup Latency Reduction
AI-assisted tooling at Lyrata
Model Outputs Processed Daily
Scalable evaluation pipelines
Onboarding Time Reduction
LyrataGPT SDK-like interface
Explore My
July 2025 – Present
Designed and deployed AI-assisted tooling to parse, validate, and surface operational knowledge from structured and unstructured sources (Python, Flask, PostgreSQL, ChromaDB, GPT-4o), reducing SOP lookup latency by ~90% and enabling consistent, auditable system behavior. Built a production-grade internal "SDK-like" interface (LyrataGPT) that delivers cited, procedure-accurate responses to technicians, improving developer-style ergonomics for non-technical users and cutting onboarding ramp time by ~50%. Applied Vector Institute DaRMoD methodologies to produce ML systems, implementing reproducible evaluation pipelines, MLflow experiment tracking, containerized services, and fault-tolerant routing under real-world conditions.
November 2022 – July 2025
Developed microservice LLM evaluation pipelines with automated rubric scoring, anomaly detection, and usage analytics, scaling to daily inferences while improving model quality by ~35%. Built SQL-backed observability and analytics systems with real-time dashboards and automated alerts, reducing debugging cycles by ~60% and improving reliability of production deployments. Led architectural refactors of distributed inference infrastructure, optimizing container orchestration, caching strategies, and load balancing to improve latency and system stability under heavy traffic. Translated business and product requirements into technical specifications for cross-functional teams, introducing standardized evaluation and release workflows that accelerated feature delivery by ~30%.
January 2022 – December 2023
Built production full-stack features using React and backend APIs, including authentication flows, SQL data models, and Stripe integrations, ensuring secure and reliable client-server communication. Worked against clearly defined API contracts, contributing to stable interface boundaries between frontend and backend systems and reducing integration errors during feature iteration. Implemented automated testing and CI/CD pipelines (Jest, React Testing Library, GitHub Actions) to support predictable releases, minimize regressions, and improve system reliability over time.
Explore My
Languages: Python, Java, JavaScript/TypeScript
AI & LLMs: GPT-4o, LLM evaluation, RAG, embeddings, semantic analysis, MLflow. I build retrieval-augmented generation systems using vector databases, design automated evaluation pipelines that process thousands of model outputs daily, and implement MLflow experiment tracking for reproducible ML workflows. My work has improved model quality by 35% through systematic evaluation and optimization.
Distributed Systems: Docker, microservices, message queues (RabbitMQ), observability, fault tolerance
APIs & SDKs: REST, gRPC, Slack API, SDK design patterns, versioning. I architect containerized microservices with fault-tolerant routing, implement observability stacks with real-time dashboards, and design SDK-like interfaces that improve developer ergonomics. My distributed automation systems integrate Slack APIs, Google Sheets, and databases with schema versioning and idempotent handlers.
Databases: PostgreSQL, MongoDB, BigQuery, vector databases (ChromaDB)
Cloud & DevOps: AWS, GCP (Pub/Sub, Airflow), CI/CD, GitHub Actions, Linux
Data & ML Systems: ETL/ELT, analytics pipelines, feature engineering, usage telemetry
Testing & QA: PyTest, Jest, automated test suites, validation pipelines. I build SQL-backed observability systems, implement comprehensive CI/CD pipelines, and develop automated testing frameworks that reduce regressions and improve system reliability.
Browse My Recent
A containerized FastAPI inference service integrating Hugging Face Transformers for real-time text sentiment prediction. Designed with Uvicorn for async concurrency and validated via Jupyter-based parallel request testing.
Built with Docker and the distilbert-base-uncased-finetuned-sst-2-english model for optimized low-latency inference.
Features async I/O, batching, and scaling via Gunicorn workers with future-ready plans for NGINX load balancing, structured logging, and Prometheus metrics.
A FastAPI microservice demonstrating end-to-end observability through Grafana, Prometheus, Tempo, and Loki. Features OpenTelemetry instrumentation for trace-to-log correlation and metric exemplars, deployed via Docker Compose.
Implements trace injection across three FastAPI services using the OpenTelemetry SDK with Prometheus metrics, Tempo tracing, and Loki log pipelines.
Load tested via Locust and k6 to validate distributed tracing and resource efficiency, with metrics visualized in Grafana dashboards for real-time debugging and performance monitoring.
A Streamlit-based analytics app for visualizing Bluecoins expense data using Python, SQL, and Docker. Designed to support intuitive budget tracking and category insights with a responsive, minimal interface.
Developed an ETL pipeline from CSV → SQL → Streamlit UI using pandas and SQLAlchemy, containerized via Docker Compose for reproducibility. Provides dynamic charts on income trends, category flow, and net balance with future integrations planned for OAuth login and hosted deployment.
A serverless financial chatbot that guides users through retirement investment planning. Built with AWS Lex for NLP dialogue management and AWS Lambda for real-time portfolio recommendations based on user risk profiles.
Implements Lambda-backed validation for user intents and secure financial calculations with algorithm-driven portfolio allocation based on risk tolerance and horizon. Built with serverless architecture for scalability and low-cost deployment, featuring an integrated conversational flow designed for accuracy, responsiveness, and ease of use.
Get in Touch
I'm always interested in discussing new opportunities, collaborations, or just having a chat about technology and innovation.
