Tarik Hasan
Software Engineer, AI · MSc Computer Science (UBC)
Software Engineer evaluating AI agents, building multi-agent workflows, and fine-tuning open-weight models to make AI systems more reliable.
What I Do
AI Evaluation
RAGAS, expert validation, rigorous benchmarks.
Agentic AI
Multi-agent workflows with reasoning and memory.
Model Fine-Tuning
QLoRA, MLflow, open-weight model benchmarking.
Production AI
FastAPI, Docker, AWS, Azure, CI/CD.
Where I've Worked
Software Engineer, AI
CanCards AI, Self-directed · Vancouver, BC
- Built and deployed CanCards AI, a live advisor answering questions about 50 Canadian credit cards with cited sources, independently delivering the full stack, from user interface to AWS cloud deployment.
- Improved search quality by combining keyword and AI search, raising relevant information found in the top eight results from 32% to 96%, measured on 100 questions reserved for final testing.
- Built a LangGraph agent that collects five required credit details before giving advice, using input validation to reduce premature recommendations from 30% to 0% across 45 test conversations.
- Automated AI testing in GitHub Actions, checking 200 search questions, 45 conversations, and 50 simulated attacks using saved AI responses, blocking software releases when results fell below quality standards.
- Cut median response time from 10.6 to 2.3 seconds by fixing a database bug found through load testing, with results measured while 15 users accessed the application at once.
Data Engineer
BC ATUS · UiTR Lab & OVI Lab, UBC Okanagan · Funded by Environment and Climate Change Canada
- Led a 7-person engineering team that shipped the GPS tracking app carried by 4,000+ participants across three study waves on iOS and Android.
- Recovered usable trip records from major data gaps by clustering participants with similar demographics and behaviour, then filling each gap with that group's most common travel pattern, weighted by time of day.
- Resolved contradictions between separately predicted trip purpose, mode, and duration using Mixed Integer Linear Programming.
- Turned raw phone logs and multi-sheet survey files into one clean research dataset with Python and SQL pipelines.
Graduate Researcher, Applied AI and Human Centric Design
OVI Lab, UBC Okanagan
- Authored SCORE, a five-dimensional framework for situated AR visualization, validated by 21 independent experts and published at ACM CHI 2026.
- Architected and shipped a collaborative mobile AR system solo over ~18 months in Unity (C#): markerless scene anchoring via ARCore / AR Foundation, real-time multi-user state sync over Photon, and a shared coordinate frame across separate rooms via Google Cloud Anchors.
- Instrumented that system to generate its own evaluation dataset: per-frame camera trajectories, response time, accuracy, and dwell duration streamed to Firebase across 576 logged trials.
- Designed and analyzed the experiment behind it: a 12-participant within-subjects study, Latin-square counterbalanced across six visualization conditions, tested with RM-ANOVA plus Friedman and Wilcoxon follow-ups to isolate the winning encoding and turn it into design guidance.
- Led a 12-person cross-functional team and weekly co-design sessions with 10 older adults, shaping an accessible VR travel app on Meta Quest 3 whose deliverables fed a grant proposal that was funded.
Graduate Teaching Assistant
COSC 341, UBC Okanagan
- Built a Python and GPT-4o pipeline using retrieval-augmented generation to review 100+ student software projects weekly against 18 criteria, completing each batch in under 40 minutes for approximately $6.
- Reduced weekly review time from 22 to 7 hours through human-in-the-loop evaluation, escalating the lowest-confidence 20% of AI reviews and validating 87% agreement with manual grading across 50 submissions.
- Debugged logic errors, build failures, and API integration issues across 500+ student codebases, translating recurring defects into evaluation criteria used by the automated code review pipeline to assess submissions.
- Co-designed the lab curriculum with three instructors, from system architecture and Figma prototyping to shipped Android builds.
Featured Work
CanCards AI
Ask a plain-English question about Canadian credit cards, get an answer with a source behind every claim. Faithfulness 0.725 → 0.851, gated by 60 CI tests.
SCORE
A five-dimensional framework for measuring how naturally AR visuals blend into the real world. Validated by 21 experts, published at ACM CHI 2026.
ScoreAI
Automates the SCORE rubric end to end: Pinecone retrieval, Gemini 2.5 Pro scoring, and schema-validated output checked against 21 expert raters.
Abu Simbel VR
An accessible VR travel app for older adults, specified by six months of weekly co-design with 10 participants before a line of it was built.
Research Papers
SCORE: A Framework for Quantifying Diegesis in Situated Visualization for Augmented Reality
Tarik Hasan, Khalad Hasan, Barrett Ens
First author
A five-dimension framework (the letters spell SCORE) for measuring how naturally AR visuals sit in the real world, validated by 21 domain experts with Krippendorff's α of 0.815–0.926 on every dimension.
Deep learning in prostate cancer diagnosis and Gleason grading in histopathology images: An extensive study
3rd of 5 authors · 145+ citations
A benchmark of deep learning approaches for reading gigapixel tissue slides across 11,000 images, spanning CNN, U-Net, and ResNet; the best architecture reached 95.3% detection accuracy. I wrote the image preprocessing and architecture comparison sections.
Academic Background
MSc in Computer Science
University of British Columbia, Okanagan
- Thesis: SCORE, a five-dimensional scoring model for how naturally an AR visualization blends into the real world, first-authored and published at ACM CHI 2026 with Khalad Hasan and Barrett Ens.
- Built the benchmark from scratch: screened 431 AR systems down to 50 papers, then scored every one into 67 structured examples.
- Validated the rubric with 21 independent AR experts, reaching strong agreement on all five dimensions (Krippendorff's α 0.815–0.926).
- Replaced guessed dimension weights with data-derived ones using PCA, and surfaced five reusable AR design archetypes through UMAP and cluster analysis.
BSc in Computer Science and Engineering
Shahjalal University of Science and Technology, Bangladesh
- Graduated with a 3.76 CGPA.
- Published a research paper on deep learning for cancer detection in an Elsevier (Q1 journal), cited 145+ times.
- Compared CNN, UNet, and ResNet models in the review, documenting reported accuracy of 89.2% for cancer grading and 95.3% for detection to summarize model performance across different tasks.
- Built and deployed FOCUS, a full-stack platform connecting students and alumni through social posts, job resources, and study materials, using MERN and Next.js as part of a two-developer team.
- Built real-time messaging with Socket.io, allowing students and alumni to send, load, and delete private messages, receive live activity notifications, and retrieve unread messages when they return online.
Leadership
Association of Bangladeshi Students, UBC
President
- Led a 250-member organization, the largest Bangladeshi student community at UBC and for most members the first place they land after moving countries.
- Ran the full year of programming end to end with the executive team: cultural events, career sessions, and sports tournaments that gave new international students a community and a professional network in their first months in Canada.
SUST ACM Student Chapter
President
- Organized 10+ webinars on software engineering with engineers from Google, Facebook, and Microsoft, covering technical interview prep, CV writing, graduate applications, and design patterns.
- Ran a Mobile Game Development training seminar in partnership with the ICT Division of the Bangladesh Government, MARS Solutions, and SHADOWHITE.
- Served as Organizing Secretary from Feb 2020 before being elected President.