Our Portfolio

A production AI platform built for a HubSpot agency that converts any public website into a working HubSpot CMS theme. The system analyzes page structure and styling with an LLM, then generates HubL templates, modules, and theme fields automatically, cutting base-theme setup that used to take 3 to 4 days of manual development down to under one hour, with a developer needing only one day for final polish. Outcome: Roughly three days saved per project; now running live in the client's production HubSpot agency workflow.

A Retrieval-Augmented Generation pipeline that interprets clinical questions and returns evidence-based procedure suggestions backed by citations from trusted medical sources. Built as decision support for healthcare professionals, not a replacement for clinical judgment. Outcome: A clinical decision-support tool that returns ranked procedure suggestions with citations and contraindication alerts. Intended to support, not replace, healthcare professional verification.

A series of LLM-assisted LinkedIn posts written to build the client's personal brand, establish thought leadership, and attract qualified leads. Posts are grounded in verifiable information and tuned to the client's voice across economics, data analytics, leadership, growth, and sustainable development. Outcome: Content that's ready to publish, matches the client's voice, and is tuned for LinkedIn engagement and lead generation.

An AI-powered virtual sales advisor that ingests scraped data from multiple car-selling platforms and recommends the best car for a user's budget, preferences, and intended use. The system understands natural-language queries, filters across price, brand, mileage, and condition, and surfaces ranked alternatives when an exact match isn't available. Supports both English and Arabic queries, and is built to scale to additional marketplaces. Outcome: A scalable recommendation agent that turns free-text user preferences into actionable, ranked car suggestions across multiple marketplaces.

CNN models trained to distinguish COVID-19 / Pneumonia chest X-rays from normal X-rays in a binary classification task. Multiple architectures and configurations were evaluated, with all code and findings published openly on GitHub. Outcome: Open-source reference implementation for COVID-19 X-ray detection, with full reproducible code and results.

A Graph Convolutional Network (GCN) for fraud detection trained on a public dataset (80/20 split). The model was attacked with adversarial samples to test how well it held up, then fine-tuned on those samples to harden it against future attacks. This also improved overall accuracy. Outcome: Produced a fraud detection model that is both more accurate and harder to fool with adversarial input.

A K-Means-based assignment engine that distributes delivery orders from multiple vendors across a fleet of riders, encoding the client's business rules into the clustering and ranking logic. Replaced a team of human dispatchers and increased throughput. Outcome: Scaled daily order assignment from ~1,000 to ~4,000 orders per day while reducing the dispatch team's workload.

A semantic matching pipeline that uses word embeddings to relate news articles across thousands of sources. A custom scraper assembled the news corpus, matches were stored in Neo4j, and an API surfaces related articles to the article a reader is currently viewing, visualised as a comparison graph. Outcome: Powered a "related news" experience that connects readers to relevant articles across the web in real time.

Convolutional Neural Network models that classify objects and digits into predefined categories and subcategories. The trained models were further fine-tuned for additional object classes, producing strong cross-domain performance. Outcome: Reusable CNN classifiers that can be fine-tuned to new object classes with minimal data and effort.

DQN and DDQN agents trained to play Flappy Bird and Lunar Lander. The agents achieve high scores and sustain long play sessions, demonstrating practical value-based deep reinforcement learning. Outcome: Working DQN/DDQN implementations that consistently outperform baselines on classic RL benchmarks.

A research study on the adversarial risks of using AI/ML for 5G network automation. The work covers supervised, unsupervised, and reinforcement-learning attack surfaces through three case studies, proposes mitigation approaches, and offers guidelines for testing how well ML models hold up to attacks in 5G contexts. Published in IEEE Internet Computing 2021. Outcome: Peer-reviewed publication that provides the 5G research community with a structured view of adversarial ML risks and mitigation strategies.

A deep reinforcement-learning algorithm (DDQN) that selects PHY-layer transmission parameters in LoRaWAN networks. Existing rule-based algorithms cause packet collisions in dense LPWAN deployments; this DRL approach reduces collisions and improves Packet Delivery Ratio by up to 500% in some scenarios. Published at IEEE LCN 2020 with a follow-up paper. Outcome: Demonstrated up to 500% PDR improvement over the best existing LoRaWAN parameter selection methods, with peer-reviewed publications in IEEE venues.

An attention-based LSTM model that generates Urdu-language captions for input images. The project required building a custom dataset of images paired with native Urdu captions, then training and deploying the model as an online service. Outcome: One of the first Urdu captioning models of its kind, ready to use online with no setup required.

A Django web application that helps researchers discover relevant scholarly work by indexing the Google Scholar feed and surfacing personalized results. NLP techniques extract key information from abstracts and keywords so that searches return contextual, not just lexical, matches. Outcome: Gave researchers a faster way to find relevant papers and collaborators, with personalized ranking based on their interests.

A modern note-taking application inspired by Notion, with built-in AI features and real-time multi-user collaboration. Users can chat with their documents, generate automatic summaries, translate content on the fly, and edit together with teammates in real time. A secure authentication layer protects it all. Outcome: A production-ready Notion alternative that combines familiar note-taking ergonomics with modern AI document intelligence.

A Python automation that searches a target website for companies, navigates to each company's document directory, and downloads every PDF. Files are hashed to detect duplicates, and NordVPN is used to rotate the bot's IP location regularly to avoid bot detection. Outcome: Fully automated PDF collection at scale, with deduplication and IP rotation that keep the pipeline reliable over long runs.

A Python script that periodically pulls the top-100 USDT trading pairs and their pricing from Binance, feeding the list into a downstream trading bot that monitors price fluctuations. Outcome: Provided a continuously refreshed list of high-volume USDT pairs that powers the client's trading algorithm.

A Python script that checks Marriott hotel pricing for locations and dates the client defines on www.marriott.com, then emails the client an HTML-formatted price comparison table for nearby hotels. Locations are managed through a MongoDB-backed admin frontend built for the workflow. Outcome: Replaced manual price checks with a scheduled report delivered by email, covering every location the client cares about.

Designed and built the backend and dashboard for a parts-catalog aggregator that ingests over 50 million product records from three high-volume supplier sites. The scraping pipeline handles each site's protections against bots and schema differences, with incremental updates feeding a unified PostgreSQL catalog and a FastAPI monitoring dashboard for extraction health and data quality. Outcome: Replaced fragmented manual extraction with a single 50M+ record catalog the client can query, monitor, and refresh, with full visibility into health for each source and data quality for each field.

A lightweight HTML/CSS/JavaScript frontend that wraps a previously delivered Python Amazon scraper. The interface lets the client add or remove product IDs and trigger scraping runs, with the new scraper extracting product data automatically. Outcome: Gave the client a usable UI on top of the scraper, replacing manual config files and making day-to-day product management much faster.

Automated data retrieval from client websites by reverse-engineering their backend APIs. Network traffic was monitored to identify endpoints, then custom Python scripts were built to fetch, clean, and process the responses. Outcome: Provided the client with continuous, reliable data extraction integrated directly into their backend workflows.

Designed and optimized MongoDB and Neo4j queries for a 14-exercise client engagement, including dataset imports, query verification, and full written documentation explaining each query and its results. Outcome: Delivered all 14 exercises on time with verified results and clear documentation, meeting the client's expectations.

Python scripts that automate data extraction from a client betting site using its underlying APIs. The scripts accept betting codes as input and return structured data, three previously-undocumented APIs were extracted from the site, and several existing code bugs were fixed. Outcome: Gave the client a fully working extraction toolchain with reliable structured output and minimal supervision required.

A real-time stock data system built around StockTitan.net's Gold Membership stream. Python scripts extract live market data (symbol, price, volume, % change, float) every few seconds, a custom dashboard displays the top 10 momentum stocks, and a module converts the latest symbol to speech. It's built for live YouTube broadcasts. Outcome: Delivered a low-latency dashboard (under 1.2 seconds end-to-end) with live audio announcements that the client now uses for real-time market broadcasts.

A Python automation that pulls leads from the client's CRM, organizes them by type and ZIP code, generates per-segment CSV files, and uploads them into Kixie. A second module extracts verified phone numbers from PlanetAltig while skipping duplicates and previously processed leads. Outcome: Eliminated manual data transfers between CRM and Kixie, with verified secondary phone numbers improving outbound contact rates.

A Python automation that identifies and processes Merchant Cash Advance (MCA) cases from public court databases. The system filters for relevant cases, enriches them with verified business contacts, and delivers daily lead reports by email. Outcome: Delivered an automated lead generation pipeline producing verified MCA business leads daily at 9 AM EST.

An automated system that scrapes court case records from US county websites, identifies commercial contract cases via keyword and category filters, and enriches them with verified mobile and email contacts using AccurateAppend and Enformion APIs. Output is delivered as clean Excel datasets for compliant SMS and email outreach. Outcome: Reduced manual case review time by over 80% and improved contact accuracy for downstream SMS campaigns.

A Python automation that submits job applications across major ATS platforms, including Workday, Greenhouse, ATS Ripple, and ASBQ Jobs. The system fills application forms, uploads resumes, and submits with minimal human intervention, with dynamic field detection per portal. Outcome: Cut manual application time dramatically and produced a working prototype across the most common ATS platforms.

A continuous scraping system that extracts car listings, pricing, and images from the major Saudi marketplaces Dubizzle.sa, Syarah, and Haraj. Listings are filtered for validity and synced into a Supabase/PostgreSQL database that powers the client's comparison platform, built with React. Outcome: Provides the client's comparison platform with real-time, validated listings from across Saudi Arabia's largest car marketplaces.

A Python trading bot for CSFloat that monitors live listings, detects underpriced skins using configurable thresholds (wear, float value, rare patterns, stickers), executes purchases, and relists items at competitive prices once trade holds expire. Outcome: Delivered a fully automated trading workflow that captures profitable listings around the clock, with configurable risk and strategy controls.

A large-scale data extraction pipeline built for Spoonful Inc. to centralize product data from major grocery retailers (Kroger, Walmart Food, Tesco, Tesco.ie, Woolworths). The scraper captures ingredients, allergens, nutrition facts, and product metadata while complying with each site's access structure. Outcome: Automated grocery data collection across multiple regions at 99% accuracy, replacing manual research and powering Spoonful's analytics and price-comparison features.

A Python scraper that collects church information, mass timings, and clergy assignments from directory.charlestondiocese.org. Parallelization is used to handle the large directory, and church images are downloaded and renamed per the client's naming convention. Outcome: Delivered a clean, ready-to-use dataset and image library for the client's diocesan directory product.

A Python scraper that extracts church information and mass timings from masstimes.org. It iterates through US ZIP codes to work around the site's missing search, runs in parallel for speed, and stores results as JSON files synced to Google Drive. Outcome: Produced a complete, structured dataset of US churches and mass times that powers the client's downstream directory product.

A high-throughput Python scraper that collects TripAdvisor listings, descriptions, and images at scale. Text is stored in PostgreSQL with categorized location tables, while images are uploaded to Dropbox. Rotating proxies enable parallel collection across many regions. Outcome: Built a continuously refreshed TripAdvisor dataset that the client uses for travel-content and location-intelligence products.

An automated Python scraper that collects wine names, providers, and prices from Wine-Searcher and identifies the cheapest provider for each wine. The scraper mimics human browsing behavior to avoid detection, and produces a fresh comparison sheet every four hours. Outcome: Delivered a reliable report of the lowest price for each wine that the client uses to drive sourcing decisions, with no manual lookups required.

A Python scraper that pulls product listings from eBay (titles, descriptions, IDs, prices, conditions) and compares them against competitor prices the client supplies. The script runs every four hours and produces a comparison spreadsheet that is shared automatically with the client. Outcome: Gave the client a continuously refreshed view of competitive pricing across eBay and partner sites without any manual collection effort.

A Python scraper that monitors selected product categories on Amazon and compares prices against the client's other sources. The script runs every four hours, exports a clean comparison spreadsheet, and delivers it to the client by email or shared cloud storage. Outcome: Replaced manual price tracking with an automated pipeline that delivers fresh price reports six times a day.

A serverless backend on AWS Lambda for an AI assistant mobile app, with DynamoDB as the primary store. Covers signup/signin, password reset, conversation history, and uploads for image, audio, and video, including a live video socket implementation. S3 handles file storage and SES handles transactional email. Outcome: A serverless backend that powers an AI mobile assistant, handling image, audio, and video uploads plus real-time video calls.

As technical lead, built the REST API backend for a complete fleet management product covering a driver mobile app, driver portal, onboarding panel, and management portal. Includes order assignment, driver onboarding, blocking/unblocking, and live location tracking, all on MongoDB and Django REST Framework. Outcome: Fleet management backend covering the full lifecycle, running in production across mobile and web surfaces.

A comprehensive Django REST Framework backend for an Order Management System, including order creation, editing, listing with advanced filtering, status management, and rider assignment. The system runs on MongoDB, queues work through Celery and RabbitMQ, and is deployed on Heroku with Slack and Sentry monitoring. Outcome: Reliable, observable order-management backend that handles the full lifecycle from creation to fleet assignment.