Harun Zafer

This project targeted legal contract analysis with natural language processing (NPL) and machine learning (ML). I architected and implemented the entire document analysis back end, which processes hundreds of contracts daily. I designed and implemented the ML pipelines to extract various information such as legal clauses, parties, effective dates, and names from contracts.

Accented characters, sometimes referred to as accents, are essential elements in written language. They frequently occur in many languages, including Spanish, French, Italian, German, and Portuguese. However, accents don't exist on the US English keyboard layout, and as a result, users tend to use the closest English version of these letters. For example, à becomes a, and ö becomes o. MagicAccents can automatically restore these letters using machine learning and natural language processing techniques with high accuracy. It currently supports 27 languages.

I've implemented this product's machine learning back end, REST API, Google Docs plugin, and desktop application using JavaFX.

Nuve is a natural language processing library for Turkish, which currently supports morphologic analysis of 35,000 words per second on an i5 2.8GHz 64-bit machine; morphologic generation; stemming; sentence segmentation and boundary detection; and N-gram extraction.

I led the design and development of WebDroid, an AI chatbot assistant that empowers websites with intelligent, context-aware interactions. Utilizing large language models (LLMs) and retrieval-augmented generation (RAG), WebDroid automates complex user queries by performing API calls and delivering precise responses.

The project’s back end leverages a serverless architecture on AWS, ensuring scalability, fault tolerance, and high availability. I implemented advanced features, including an event-driven web scraping service, and optimized LLM parameters to improve response quality. I developed a user-friendly panel on the front end with SvelteKit and DaisyUI, incorporating secure authentication with email verification and OAuth.

I also led a team of four engineers, driving efficient workflows with CI/CD pipelines deployed via GitHub Actions to Vercel. Additionally, I built the company and product websites with Svelte and Tailwind CSS, ensuring seamless user experiences. WebDroid exemplifies cutting-edge AI integration with robust engineering to transform customer engagement for businesses.

I designed and implemented a next-generation AI agent to empower sales teams with advanced content processing and retrieval capabilities. The project focused on enhancing retrieval-augmented generation (RAG) systems to deliver precise and actionable insights from vast datasets.

Key contributions included building a scalable content processing pipeline using Azure Functions to handle periodic and on-demand NLP tasks. I developed a range of AI services, such as entity extraction, deduplication, document categorization, summarization, and capability detection for client content. The back-end infrastructure was optimized for performance and scalability with Azure Functions and PostgreSQL supporting custom user workflows and analytics.

To ensure robust quality, I designed retry mechanisms for AI-generated responses and improved schema designs for efficient database operations. Additionally, I researched and proposed knowledge graph integrations to extend the system’s capabilities.

This project demonstrates expertise in delivering scalable AI-driven solutions, combining cutting-edge NLP techniques with seamless integration into cloud platforms.