Md Ashikur Rahman

• Initiated CogniX as an early-stage R&D effort to explore generalized multi-reference image fusion and prompt-guided image editing.
• Exploring architecture options for generating a coherent output image from a prompt and 1-3 reference images.
• Evaluating Qwen-Image-Edit-2509 and related multi-image conditioning approaches for tasks such as product visualization, garment replacement, scene composition, style transfer, and object-level image editing.
• Experimenting with LoRA/QLoRA fine-tuning strategies, 4-bit NF4 quantization, and PyTorch/Diffusers-based training workflows.
• Investigating dataset design for variable-length multi-reference inputs, semantic instructions, attribute-rich prompts, and preservation constraints.
• Testing ComfyUI-based inference workflows for multi-image upload, graph-based execution, VRAM control, and LoRA-enabled generation.

• Conceived the product idea and co-led the system design for a unified cross-brand virtual try-on platform.
• Designed the user workflow where shoppers can browse products from hundreds of brands, virtually try them on, and purchase from one place.
• Co-architected the platform architecture and technical pipeline, including a Dockerized FastAPI backend behind Nginx, SQL Server data layer, Google Cloud Storage for assets, Redis for caching, recommendation design, and 2D virtual try-on workflows.
• Fine-tuned CogniX, an internal Qwen-based LLM using LoRA, for garment understanding, catalog matching, product recommendation, and product-to-user try-on reasoning across diverse apparel categories.

• Leading development of an AI site-audit platform for a major US retail client, automating quality checks across product imagery, catalog data, and backend content health.
• Built image-audit workflows to detect inappropriate content, sheer-garment visibility, offensive tattoos, vulgar copy, crop inconsistencies, retouching defects, face-crop issues, and duplicate or near-duplicate images.
• Developed data-audit capabilities to validate product copy, color and style attribution, product-type metadata, image origination tags, and DAM/CMS integration requirements.
• Designed backend audit checks for inventory gaps, broken image and swatch links, alt-text quality, and pages requiring updates or removal.

• Co-participated in the system design and architecture of Omnimage.ai, an AI platform for generating images and videos from text prompts and reference images.
• Helped design the technical pipeline for prompt-based generation, reference-image conditioning, image editing, video generation, background processing, asset handling, and downstream platform integration.
• Enabled users to select from multiple state-of-the-art generation models, including Gemini Nano Banana, Kimi Krea, and other image/video generation models.
• Built production APIs and backend workflows for generation requests, background processing, and delivery to downstream systems.
• Supported onboarding and usage across 200+ brands using the platform for creative and product-image workflows.

• Built a 3D asset pipeline that converts 360° video, multi-view image sequences, and reference images into textured 3D assets.
• Implemented camera pose estimation, structure-from-motion, and multi-view geometry workflows for automated 3D reconstruction.
• Explored image-to-3D generation using sparse 3D latent representations, transformer-based generation, and multi-format decoding.
• Optimized reconstruction quality and processing efficiency to reduce manual 3D cleanup and editing effort.

• Developed and optimized deep-learning models for salient object detection and image segmentation in automated background-removal workflows.
• Improved segmentation quality and processing speed over the previous production baseline, achieving a 17% accuracy gain and 30% faster processing.
• Optimized inference for large image inputs, processing 257 MB images in 2.27 seconds using U-Net/RSU-style architecture blocks and advanced pooling.
• Built the production segmentation pipeline for large-scale background-removal workloads, supporting 4.5M+ images processed globally across hundreds of customers worldwide.

Production client: leading European luxury resale marketplace

• Helped optimize a .NET and SQL Server workflow platform for high-volume image-processing operations.
• Improved backend flows across FTP ingestion, order management, editor assignment, AI processing, QC validation, and automated delivery.
• Worked on job routing, batch processing, status tracking, and handoff automation to reduce manual work across the production pipeline.
• Supported integration of Omnimage.ai and Retouched.ai for AI-assisted background removal, image enhancement, and asset preparation.
• Helped the team support production throughput of about 17K images per day, with average batch turnaround of about 65 minutes.

• Built an audio intelligence workflow for customer-service QA using speech transcription and LLM-based phrase extraction.
• Deployed SeamlessM4T for audio-to-text transcription across customer-service conversations.
• Integrated Llama 2 to extract contextual key phrases from transcribed calls for QA review and downstream analysis.
• Built API workflows for transcription, phrase extraction, result validation, and integration with QA systems.
• Reduced manual review effort by automating transcription and key-phrase spotting across customer-service audio.