Leading frontend development for production AI applications, with responsibilities spanning backend performance, automated testing, and cloud delivery.
Fullstack AI developer · Toronto
Hey! I’m Nikolas.
I’m a fullstack AI developer at Intact, focused on building AI-powered products that are useful, reliable, and easy to use.
I work across the stack, from TypeScript and React interfaces to Python services and cloud infrastructure.
Experience
Designed cross-platform automation and document-data tools that removed more than 500 hours of manual work each week for internal teams.
Selected work
Production experience. Research depth.
Engineering work at Intact, followed by academic projects in machine learning and model evaluation.
Professional experience
Production AI engineering at Intact
At Intact Insurance's Data Lab, I lead frontend development for production AI applications, with responsibilities spanning backend performance, automated testing, and cloud delivery.
- Frontend ownership
Lead React and TypeScript development, extending shared application code for additional teams.
- Backend modernization
Migrated a FastAPI/PostgreSQL backend from synchronous to asynchronous I/O across database access, downstream-service clients, and API endpoints.
- Testing and delivery
Write automated tests with Vitest and pytest, maintain GitLab CI/CD pipelines, and contribute to AWS infrastructure through Terraform.
02
Undergraduate thesis · 2024
Distilling Chess Intelligence
I built an end-to-end pipeline that generated 15,000 Leela Chess Zero evaluations and compared nine smaller model architectures. Deeper models fit the training data more closely, but none beat a linear baseline on held-out positions. It was a useful lesson in generalization, representation, and experimental scope.
Position
03
Engineering capstone · 2023
Intracranial Hemorrhage Detection
As part of a five-person capstone team working with Trillium Health Partners, I focused on CT-image preprocessing and model evaluation. I worked on preparing 500,000+ DICOM images using pydicom and CT windowing, then analyzed predictions across five hemorrhage subtypes and overall detection on 78,545 test images. The team's baseline achieved an F1 score of 0.852 for detecting any hemorrhage, with substantial variation across subtypes.
Read the case study: Intracranial Hemorrhage DetectionTechnical profile
From model behavior to product behavior.
AI & machine learning
From experimental models to production retrieval and language-model workflows.
- PyTorch
- TensorFlow
- scikit-learn
- RAG
- Computer vision
Product engineering
Interfaces and services that turn model capabilities into dependable user experiences.
- TypeScript
- React
- FastAPI
- PostgreSQL
- Vitest & pytest
Cloud & infrastructure
Repeatable systems for deploying, operating, and improving intelligent products.
- AWS
- GCP
- Terraform
- Docker
- GitLab CI/CD
- Linux
About
Curious about the whole system.
I’m a Toronto-based developer who likes working where machine learning meets product engineering. My background in Engineering Science shaped how I approach the work: start with the underlying question, build the full path from data to interface, and be honest about what the evidence does and does not show.
Let’s connect
Building something intelligent?
I’m always interested in thoughtful engineering problems and teams building useful AI products.
niko.marinkovich@gmail.com