I engineer AI digital operating systems that turn manual industries into AI-driven markets.

I use Full-Stack Engineering to bridge the critical gap between Customer Demand and Operational Execution. I do it with governance, scalability, and ethics built in.

  • The Infrastructure: I architect multi-tenant, asset-light systems that scale without overhead.
  • The Connective Tissue: I deploy NLP, Computer Vision, and Predictive Models (Regression, Classification, Time-Series) to automate decisions.
  • The Outcome: I build platforms that autonomously orchestrate logistics, secure revenue streams, and dominate markets.

Flagship Platform: MESS Tracker

The Asset-Light Operating System for Waste Management

MESS Tracker is a multi-tenant SaaS marketplace that digitizes the entire waste service lifecycle. It serves as a centralized operating system, replacing fragmented, manual workflows with an intelligent, data-driven platform that connects demand, dispatch, and execution.

System Architecture: The "Three-Gate" Model

🔐 Zero-Trust Security

Implemented a Three-Gate Architecture to strictly isolate:
• Public Users (Magic Link Auth)
• Internal Ops (2FA Fortress)
• Drivers (Device-Bound Tokens)

💾 The "Jukebox" Data Model

Designed a hierarchical relational model separating Service Definitions from Tenant Availability.

The system orchestrates 98+ micro-services and 18+ core clusters across 8+ tenants , utilizing Redis to serve 10,000+ cached keys with millisecond latency.

Python (Django) PostgreSQL (Multi-tenant) Redis (Broker) Celery (Async) Docker
MESS Tracker End-to-End Architecture
Figure 1: The "End-to-End" Orchestration Flow: Connecting Demand Creation to Settlement.

AI as Infrastructure

Operational Intelligence (Not a Feature)

AI inside MESS Tracker is an operational infrastructure embedded directly into the "Three-Gate" workflow. It coordinates Vision, Language (Multilingual), Risk-analysis (Regression/Classification), and Time-Series intelligence to automate logistics.

MESS Tracker AI Pipeline
Figure 2: The ML Pipeline. Data ingestion (ETL), Multi-model training (RoBERTa, YOLOv11, LSTM), and Cloud Deployment.

Implemented Capabilities

💬 NLP & Transformers

Intent Recognition: Acts as the "Front Door," routing unstructured requests to DB tables.

RoBERTa SpaCy GPT-2

👁️ Computer Vision

Automated Inspection: Real-time waste classification and contamination detection.

YOLOv11 Segmentation OpenCV

📈 Predictive Ops

Demand Forecasting: Predicting operational load to optimize driver scheduling.

SARIMA LSTM / GRU

🎨 Frontend & Execution

From Insight to Automation

The User Interface is the bridge between the AI models and the real world. Below is the operational workflow: Dashboarding → Predictive Planning → AI Automation → Final Dispatch.

👆 Tap any slide below to view full screen

User Dashboard Overview
1. User Command Center
Service shortcuts and a Multilingual AI Chatbot.
Analytics
2. Live Analytics & Forecasts
7/14/30-day demand forecasting models.
Smart Calendar
3. Smart Scheduling Calendar
Risk Analysis based scheduling.
4. Demo: Predictive Scheduling
Processing structured data to suggest slots.
5. Demo: NLP Scheduling
Converting text requests into events.
6. Demo: Computer Vision
Automated waste identification.
7. Demo: Dispatch Execution
Routing the finalized task.

(Swipe right to view AI Demos & Dispatch Video) →

🏆 Real-World Impact

Beyond Localhost: Defense & Recognition

MESS Tracker wasn't just a theoretical exercise. It was a rigorous Capstone project that involved academic defense, stakeholder presentations, and public showcasing. It stands as a proof-of-concept for how AI can tangibly modernize municipal infrastructure.

Capstone Project Defense
1. The Technical Defense

Presenting the "Three-Gate" Architecture and AI Integration strategy to the review board.

Capstone Showcase with Supervisor
2. Industry Showcase

Demonstrating the live MVP alongside Project Supervisor, Reeta Suman, at the Capstone Exhibition.

📜 Official Project Poster

The comprehensive breakdown of the CRISP-DM Methodology used to build MESS Tracker.

  • Methodology: Phase 1-5 (Business Understanding to Maintenance)
  • Tech Stack: Django, YOLOv11, LSTM, SARIMA
  • Outcome: A scalable, AI-driven operating system.
View Full Size Poster →
MESS Tracker Official Poster

Collaborative AI: StrokeRisk System

Clinical Decision Support & MLOps Governance

The Leadership Context: While MESS Tracker showcases my solo architectural skills, StrokeRisk demonstrates my ability to lead and integrate within high-performance teams.

Leading Group 4 (G4 Pulse): Fuad, Preston and Marrium in our first semester for development and Group 2 (G2): Kevin and Shalin in our second semester for MLOps, I orchestrated the transition from a raw dataset to a governed, FDA-aligned deployment. We moved beyond "just coding" to building a compliant, auditable lifecycle.

StrokeRisk is an end-to-end clinical decision-support system built using the CRISP-DM Methodology. It leverages a Soft-Voting Ensemble Model to predict stroke probability with high recall, ensuring high-risk patients are identified early.

⚙️ Phase 1: The Ensemble Innovation

Phase 1 Development Team

To tackle the "Accuracy Paradox" in medical AI (where 95% accuracy hides missed diagnoses), we rejected single models in favor of a Soft-Voting Ensemble.

  • Data Balancing: Applied SMOTE to correct the 4.9% minority class imbalance, achieving a perfect 50/50 training split.
  • The Architectures: Aggregated Random Forest, XGBoost, and Extra Trees.
  • The Result: The ensemble stabilized variance and maximized Recall (Safety).
PyCaret SMOTE XGBoost

🛡️ Phase 2: Governance-as-Code

Phase 2 MLOps Team

We didn't just train a model; we built an immutable audit trail using MLflow to satisfy PIPEDA & FDA SaMD reproducibility guidelines.

  • Reproducibility: Enforced `conda.yaml` environment locking to prevent dependency drift.
  • Auditability: Every run logged Git Hashes, Dataset Digests, and Parameters.
  • Gated Promotion: Implemented a strict Staging → Production workflow requiring governance approval.
MLflow CI/CD Audit Logs

🏆 Validated Performance

95.8%
Accuracy
0.992
AUC Score
96.5%
Recall (Safety)

Cloud Deployment

The final system was deployed on Streamlit Cloud, serving the MLflow-registered model via a REST API. The frontend was designed using Human-Centered Design principles (Fogg Behavior Model) to ensure clinician trust.

Features Implemented:
• Real-time Risk Assessment
• SHAP-based Explainability (Why did the model say yes?)
• PDF Report Generation

Figure 3: Live Demo: The Clinician Dashboard & Explainable AI

Ready to build for the real world?

I am currently based in Calgary, AB and available for
Product Engineering & Management roles.

📬 Send me an Email in Connect on LinkedIn