Zubair – Python, OpenAI, LangChain

An AI-focused financial technology project centered on developing LLM, RAG, and multi-agent workflows for transaction-related use cases. The solution combined vector-based retrieval, scalable cloud infrastructure, backend API services, and modern web interfaces to support intelligent data processing and user-facing workflows.

• Designed and implemented RAG workflows using PostgreSQL and pgvector to improve retrieval accuracy for complex transaction analysis;
• Developed and deployed multi-agent LLM systems for production transaction-focused use cases, leveraging Azure and Argo CD;
• Built scalable FastAPI services to support AI-powered workflows and backend integrations;
• Designed React-based interfaces that enabled users to review, validate, and act on AI-generated outputs;
• Collaborated across AI, backend, and frontend components to deliver reliable, production-ready solutions.

An AI-powered resume and portfolio platform that helps users create polished, publish-ready career materials in under 5 minutes through AI chat, voice-guided assistance, and intuitive editing controls.

• Built a guided AI chat and Voice AI Agent experience that accelerated resume and portfolio creation without requiring design expertise;
• Designed the product vision for AI-generated initial designs with full user control over editing, personalization, and final presentation;
• Created a streamlined workflow that reduced the time from setup to publish-ready resume and portfolio content to under 5 minutes.

A large-scale design platform serving over 40 million users, focused on providing accessible and intuitive tools for creating, editing, and managing digital content. The project involved designing and improving user experiences across a high-traffic product ecosystem with a strong focus on usability, scalability, and user engagement.

• Shipped AI image tools, including Upscaler, Deblurrer, Object Remover, and Object Replacer, improving visual content creation for 40M+ users;
• Built an AI Content Resizer to adapt designs across formats with high layout accuracy and reduce manual resizing work;
• Fine-tuned multimodal LLM workflows powering AI Designer and Edit with AI to improve generation quality and editing speed;
• Improved template recommendation and image selection through multimodal ranking for more relevant design suggestions;
• Developed an interactive chatbot that engaged users in a conversation, understood their requirements, and identified the best-matching template;
• Introduced the dynamic generation of customized templates, where the chatbot curated content, images, and icons based on the user’s description;
• Refined Visme’s Image Search algorithm to provide more accurate and contextually relevant results.

An end-to-end computer vision pipeline for automated basketball highlight generation and match statistics from recorded video.

• Delivered automated basketball highlight generation from raw match video, reducing manual editing effort for sports content workflows;
• Optimized model inference with TensorRT to speed up video processing and improve pipeline efficiency;
• Created visual analytics to review highlights, player actions, and match statistics in a more structured workflow;
• Researched and pipelined state-of-the-art computer vision models needed for separate modules;
• Converted and optimized the models to TensorRT and achieved 18 fps on average using only a single V100 GPU;
• Developed advanced graphic visualization using OpenCV that provided reasoning behind the clipped highlight video and helped analyze false positives.

An AI-enabled note-taking application with offline multilingual support.

• Built NLP features for text classification, entity recognition, and contextual translation to improve note understanding;
• Implemented Transformer-based offline multilingual support using BERT-style models for more accessible note-taking workflows;
• Developed contextual word-to-word translation features that improved multilingual usability inside the product;
• Implemented an NLP engine that was trainable on demand based on customer datasets annotated using an easy-to-use GUI;
• Built lambda-based functions for text classification, entity recognition, question and answer generation, and PDF text extraction.