Ibrahim – Python, LLM evaluation, LLM benchmarks

An AI-powered product visualization platform that lets users preview how a product would look in a real-world setting by placing it onto their own photos before purchase. The product generates realistic previews with controlled product placement and automated image preparation, helping users evaluate products beyond standard catalog images.

• Designed and built the end-to-end AI image generation pipeline for product visualization using Imagen 4 via fal.ai;
• Owned input processing, including background removal, sticker/marker detection, and perspective correction for user-submitted photos;
• Scoped and planned the deployment architecture, including Railway for backend hosting and Cloudflare Pages for frontend;
• Engineered a deterministic color marker system using lime green/red markers to provide precise spatial control over product placement;
• Improved output consistency by feeding pre-arranged reference images directly into the generation pipeline;
• Reduced hallucination by anchoring generated results to predefined visual references;
• Took the project from a raw AI generation concept to a defined, production-ready pipeline;
• Identified and solved input cleaning bottlenecks that are often overlooked in generic AI demonstrations.

Developed a computer vision system for automatically detecting agricultural field boundaries from satellite imagery, replacing a significant amount of manual field delineation work. I worked across the ML lifecycle, from satellite data preparation and model training to production deployment and monitoring.

• Designed and developed the ML pipeline for field-boundary segmentation from satellite imagery;
• Built data preprocessing and image-tile generation workflows for large-scale satellite data;
• Trained and evaluated segmentation models and improved model performance through preprocessing, augmentation and feature engineering;
• Owned model deployment and production monitoring;
• Implemented evaluation using IoU, Dice score, and under/over-segmentation metrics;
• Worked with geographically diverse datasets and investigated regional differences in model performance;
• Helped reduce manual labeling effort and improve field-data accuracy by around 40%.

Developed a GenAI-powered question-answering system for an agricultural data platform. The system allowed scientists, agronomists and other users to interact with both structured and unstructured agricultural data using natural language.

• Designed and developed RAG-based workflows for querying unstructured agricultural knowledge;
• Implemented natural-language-to-SQL capabilities for structured data;
• Combined vector and keyword search to improve information retrieval;
• Designed routing between smaller and larger language models based on query complexity and cost;
• Worked on evaluation strategies for retrieval and generated answers, including representative question sets and golden datasets;
• Focused on reliability, hallucination reduction, context quality, and production cost;
• Collaborated with data scientists, engineers, and domain experts to improve the system.

Developed machine learning solutions for detecting fraudulent transactions in an e-commerce environment. The project focused on identifying suspicious transactions while controlling false positives and handling delayed fraud labels.

• Developed and evaluated machine learning models for transaction fraud detection;
• Designed features based on transaction, customer and behavioral data;
• Worked with delayed labels caused by chargebacks, disputes and investigation processes;
• Evaluated models using precision, recall and F1 while balancing fraud detection against false positives;
• Built data processing and ML pipelines for large-scale transaction data;
• Worked on production deployment, monitoring and model iteration;
• Collaborated with engineering and business teams to translate fraud risks into ML solutions.