Hi! I'm Sirjan

Context

In a previous role, ops teams received a daily agent performance report I built. It was used to make real decisions about which agents needed coaching, which showed no signs of improvement, and which were performing well enough to benchmark others against. This project is a public recreation of that architecture using a synthetic dataset I designed to mirror real collections industry behavior, because the real client data stays private.

Action

• Data Engineering: Used the Faker library in Python to generate a dataset with realistic distributions, including time-of-day call patterns and holiday-driven behavior, to avoid the statistical flatness of purely random data.
• Business Logic: Programmed agent-level performance constraints that reflect how real collections KPIs behave across RPC rates, payment conversions, and silence time, so the data could support meaningful drill-down analysis.
• Visualization: Designed a Tableau dashboard with full drill-down capability, built to the same standard as a production reporting environment.

Context

Algonquin College's Applied Research department was managing faculty project registrations and external partner matching entirely through email and spreadsheets. There was no centralized way to track submissions, match faculty to partners systematically, or give the administrative team any consistent visibility into program activity. Built as a capstone project with the Applied Research department as the client, the goal was to replace that fragmented process with a single automated platform.

Action

• Centralized Hub: Built role-specific SharePoint pages for faculty, external partners, and administrators, each with tailored forms, access permissions, and partner availability lists to standardize how data entered the system.
• Smart Automation: Designed Power Automate workflows to handle form submissions end-to-end, including automated faculty-partner matching by academic school and term, conditional email logic based on contact disclosure preferences, and real-time notifications.
• Executive Visibility: Built a four-view Power BI dashboard covering faculty participation, project trends, external partner engagement, and an equity tracking view for Indigenous participation, powered directly from SharePoint lists.

Context

This project started during my undergrad as a basic CNN-based image captioning experiment with a local UI and limited model quality. I rebuilt it from scratch at Algonquin once I had access to better tools and infrastructure. The goal was the same, generating accurate natural language descriptions from images, but this time done properly: a fine-tuned BLIP model, a real dataset, formal evaluation, and a deployable application.

Action

• Model Fine-Tuning: Fine-tuned Salesforce's BLIP model on 20,000 images from the Flickr30k dataset using mixed-precision training and gradient accumulation to work within GPU constraints, with the remaining 11,784 images held as an unseen test set.
• Performance Evaluation: Tracked model quality using BLEU score across three stages: zero-shot baseline, a 5,000-image subset, and the final 20,000-image run. Each iteration measurably improved the model's ability to describe entities, actions, and scene context.
• Deployment: Built a Streamlit application with session-based authentication, replacing the original local Tkinter prototype with a clean, shareable interface structured for future cloud hosting once model size constraints are resolved.

Context

Graduate school applications involve a lot of uncertainty with limited objective feedback. Built during my graduate studies in BI Systems Infrastructure at Algonquin, this project uses a UCLA admissions dataset to explore whether standardized academic inputs like GRE, TOEFL, and CGPA can reliably predict admission probability. The goal was to practice building end-to-end ML pipelines with a clean, usable interface.

Action

• Neural Network Architecture: Engineered a Keras regression model to process multi dimensional inputs including GRE, TOEFL, and CGPA scores.
• Streamlit Integration: Developed a web application that translates complex probability scores into intuitive categories like High Chance or Low Chance.
• Reproducible Research: Implemented a robust data pipeline and random seeding to ensure consistent, reliable predictions with an R squared score of 0.74.

Context

Property pricing is driven by a mix of physical attributes and broader market conditions that are hard to weigh manually. Built during my graduate studies at Algonquin, this project uses a public real estate dataset to explore how ensemble models handle non-linear relationships between features like lot size, property type, and economic indicators including recession periods and neighborhood popularity. The goal was to practice feature engineering and model deployment end-to-end.

Action

• Ensemble Learning: Implemented a Random Forest regression model to handle the complex, non linear relationships between property tax, insurance, and sale prices.
• Feature Engineering: Incorporated economic indicators such as "Recession" and "Popularity" to refine predictions beyond standard physical dimensions like bedrooms and lot size.
• Interactive UI: Developed a Streamlit application with intuitive input forms and tooltips to ensure the model remains accessible and transparent for non technical users.

Context

Most business reporting stops at the numbers. This project goes further. It takes a structured sales dataset, calculates performance metrics across category, region, segment, and sub-category dimensions, and feeds those metrics into a prompt engineered LLM pipeline that produces an insight driven executive narrative with a traffic light scorecard, an executive brief, and collapsible detailed analysis sections.

Action

• Data Grounding: Pre-calculated all metrics in Python and passed verified numbers with direction labels to the LLM, eliminating a whole class of hallucination errors
• Prompt Engineering: Designed and iterated five modular prompt files across nine versions to produce C-suite quality narrative output
• Report Generation: Built an interactive single file HTML report with collapsible sections, color coded margin delta tables, and a performance scorecard