Computer Science undergraduate at Canadian University Dubai (GPA: 3.97) with hands-on experience in data engineering, full-stack development, machine learning, computer vision, and LLM-powered applications.

I build end-to-end systems — from orchestrated cloud data pipelines and deployed multi-user web services to object detection models — and I'm comfortable across the whole workflow: infrastructure as code, ELT modelling, API design, controlled experimentation, and quantitative evaluation.

📫 LinkedIn • Email

Languages: Python • TypeScript • JavaScript • Java • SQL • C++ • Prolog • Lisp

Data Engineering: Apache Airflow • dbt • BigQuery • Google Cloud Storage • Terraform • Docker Compose • Elementary • Looker Studio

ML & Computer Vision: YOLOv8 • PyTorch • scikit-learn • pandas • NumPy • Transfer Learning • Object Detection

AI & Automation: Gemini • Groq (Llama 3.3) • Ollama (DeepSeek-R1) • FAISS • Browser Automation

Frameworks & Tools: FastAPI • React • Tailwind CSS • Streamlit • Git • GitHub Actions • pytest • Playwright • Roboflow • Google Colab

Databases: PostgreSQL • SQLAlchemy • BigQuery • SQLite

Cloud & Deployment: Google Cloud Platform • Vercel • Render • Neon

Python Airflow dbt BigQuery Terraform Docker Looker Studio

Batch data pipeline ingesting air-quality data from the OpenAQ v3 API, comparing the UAE and Pakistan on PM2.5 against WHO 2021 guidelines.

• Backfilled 1.8M measurements in resumable 60-day chunks, each count-reconciled against the API before it is checkpointed
• An Airflow DAG fans out one dynamically-mapped task per sensor behind a rate-limit pool; a Dataset event then schedules a dbt transform DAG that astronomer-cosmos renders as one Airflow task per model
• staging → daily aggregates → WHO-exceedance marts across 121 dbt nodes and 83 data tests, with every GCP resource provisioned by Terraform and five required CI checks on main
• The finding: Pakistan's PM2.5 exceeded the WHO 24-hour guideline on 100% of 455 shared days (mean 75 µg/m³) against the UAE's 94% (mean 37 µg/m³) — published alongside the coverage gap that qualifies it, ~180 reporting stations a day versus ~8

React 19 TypeScript FastAPI PostgreSQL SQLAlchemy Gemini PWA

A deployed multi-user nutrition tracking app, rebuilt from a single-user Streamlit prototype into a decoupled frontend/backend architecture. 21 shipped phases · 725 tests · 50 API endpoints.

• Multimodal AI meal analysis (Gemini) — describe a meal, record a voice note, or photograph it; any one produces a structured estimate and combining them sharpens it. Results are calorie and macro ranges with per-ingredient confidence, never a false-precision single number
• The model's accuracy is measured, not assumed. Every estimate is logged and compared against what the user actually saved — interval coverage and directional bias per macro, with sample sizes — and it refuses to report where the evidence is too thin
• It uses your saved foods instead of guessing them: the AI estimates the portion, your library supplies the macros — splitting the work along the line each side is actually good at
• Depth beyond logging: a daily burn measured from energy balance rather than a formula, calorie banking across days, an eight-check weekly review where every figure carries its own window and sample size, water/step/supplement trackers, and meal sharing by self-contained code
• Argon2id + JWT auth with per-IP rate limiting; cross-tenant isolation proved twice — 49 in-process tests and a 127-check smoke script fired at a running server
• Installable PWA, five-gate CI, Alembic migrations on every deploy, on Vercel + Render + Neon

Full feature list, architecture and screenshots in the project README.

Python YOLOv8 Roboflow PyTorch Google Colab

Trained a YOLOv8s object detection model on a custom 2,069-image campus dataset to detect fire alarms, wet floor signs, emergency exits, and safety helmets in real time.

• Achieved 96.8% mAP@0.50 after diagnosing class imbalance and consolidating exit sub-classes (+15.8 pp improvement)
• Ran 3 controlled improvement experiments; hyperparameter tuning was the best strategy (94.52% mAP on test set)
• End-to-end pipeline: Roboflow annotation → COCO transfer learning → 10.4ms/frame inference

Python Standard Library HTTP Windows

One place to see and launch every game on a Windows PC — Steam, Epic, Xbox, Riot, and the loose folders that belong to no launcher at all — with real cover art for all of them.

• Pure Python 3 standard library: no pip install, no npm, no build step, and nothing in the cloud
• Runs a small HTTP server on 127.0.0.1 and opens a browser window at it; launcher detection, executable heuristics and icon extraction are all hand-rolled

Python Pandas scikit-learn

End-to-end ETL pipeline covering ingestion, profiling, cleaning, feature engineering, and classification modeling on large-scale taxi trip data — achieving 97% prediction accuracy.

Python Streamlit Gemini LLM Ollama DeepSeek-R1 Browser Automation

AI-powered browser automation tool that logs into the CUD student portal, scrapes course offerings, and enables natural language queries via integrated LLMs.

SQL

Multi-dataset SQL analysis exploring correlations between community hardship, crime rates, and school performance — including views, stored procedures, and transaction management.

Python Streamlit Gemini LLM APScheduler SWI-Prolog

Natural language medication scheduling app using Gemini LLM for input parsing, APScheduler for automated reminders, and Prolog logic rules for drug interaction safety checks.

Python Linux systemd nginx Tailscale

A PaperMC Minecraft server and a custom web control panel running on a headless Ubuntu mini PC at home, for a handful of friends.

• PaperMC under systemd, with a Python control panel served over HTTPS behind nginx — start/stop, live console, player management, plugin installs and metrics from a browser
• Tailscale for remote administration, a tunnel for friends to connect through, and scripted backups with off-box copies

No public repository — the repo is a private server backup rather than a portfolio project.

Python FAISS sentence-transformers Groq FastAPI

A research assistant over a local corpus of AI/ML papers. An agent decides, per query, whether to answer from the papers (dense retrieval over a FAISS index) or from a live web search — the routing decision is the point of the project.

Deliberately no RAG frameworks — no LangChain, no LlamaIndex. The chunker, the embedding and indexing step, retrieval, the ReAct loop and tool dispatch are all written directly; libraries are used only for primitives. Every model call goes through one entry point, so swapping Llama 3.3 70B on Groq for a local Ollama server is a one-file change.

Status: the PDF ingestion and text-cleaning layer is built. Chunking, the FAISS index, retrieval, the agent loop, the evaluation harness and the API are still to come.

Interactive browser-based tools built for BCS 309 – Algorithms I. No dependencies — pure HTML, CSS, and vanilla JavaScript, deployed on GitHub Pages.

HTML CSS Vanilla JavaScript

Six-section interactive lecture covering brute force O(n·2ⁿ), dynamic programming O(n·t), pseudo-polynomial complexity, and a formal NP-completeness proof via 3-SAT reduction. Three visualization modes: DP table animation, brute force bitmask walkthrough, and side-by-side comparison with live operation counters.

HTML CSS Vanilla JavaScript

Step-by-step visualization of the Huffman coding algorithm — min-heap construction, greedy merges, tree building, and code generation. Includes encode/decode, complexity comparison charts (Huffman vs Shannon-Fano vs Arithmetic), and an exchange-argument proof of greedy optimality.

• Introduction to Modern AI — Cisco Networking Academy (Apr 2026)
• Databases and SQL for Data Science with Python — IBM, Coursera (Jul 2024)
• C Essentials 1 — Cisco Networking Academy (Apr 2026)
• What is Data Science? — IBM, Coursera (Dec 2023)