Md Ashikur Rahman

• Started and architected Cognix, an ongoing generalized image fusion and editing system that takes a prompt and 1-3 reference images to generate a single coherent output image.
• Designed the system for prompt-guided multi-image conditioning tasks such as product visualization, identity-preserving edits, scene composition, style transfer, garment replacement, and object-level image manipulation.
• Building the model strategy around Qwen-Image-Edit-2509, using its native multi-image input, Qwen2.5-VL encoder for image understanding, MMDiT transformer for image generation and fusion, and frozen VAE for decoding.
• Implemented a config-driven fine-tuning pipeline using QLoRA with 4-bit NF4 quantization, enabling a ~20B model to train on a 24 GB RTX 4090, with bf16/H100 support through configuration changes.
• Developed variable-length multi-reference dataset support for 1, 2, and 3 reference images, using semantic instructions, attribute-rich prompts, and preservation clauses.
• Built training, validation, evaluation, and inference workflows, including dataset validation, model loading, LoRA training, checkpoint evaluation, and prompt-based generation.
• Verified end-to-end training on an RTX 4090 with a successful micro-train run, 94M trainable LoRA parameters, and checkpoint generation.
• Integrated inference through ComfyUI using FP8 model execution, API-based graph submission, multi-image upload, VRAM controls, and LoRA-enabled generation.
• Currently evaluating base-model behavior across multi-reference editing tasks before building the full fine-tuning dataset.

• Conceived and led the end-to-end development of a unified virtual try-on platform where shoppers can search across brands, try products on with AI, and purchase from one place.
• Fine-tuned Qwen with LoRA for garment understanding, catalog matching, and product recommendation across diverse apparel categories.
• Built AI-powered 2D virtual try-on workflows for apparel such as t-shirts, jeans, dresses, shoes, and sarees.
• Onboarded 170+ brands, including major international names, into the platform’s searchable product catalog.
• Reached a 16% purchase conversion rate, reflecting strong shopper confidence from try-before-buy experiences.

• Led development of Omnimage.ai, an AI-powered image generation and editing engine for scalable product-image workflows.
• Fine-tuned LLaMA-based components to support high-volume image-processing requests with 10K+ daily requests and sub-18-second inference latency.
• Built production inference APIs for image editing, background processing, and integration with downstream platforms.
• Integrated Omnimage.ai into automated image-processing workflows, including background removal and enhancement tasks used by enterprise e-commerce operations.

• Led development of an AI-powered catalog intelligence platform for product understanding, search, and structured attribute extraction.
• Built workflows for product classification, metadata enrichment, and semantic matching across large e-commerce catalogs.
• Applied computer vision and language-model pipelines to improve product discovery, catalog quality, and downstream recommendation workflows.
• Designed API-driven services for integration with product catalogs, content systems, and client-facing applications.

• Built a reconstruction pipeline that converts 360° video and multi-view image sequences into high-quality 3D meshes in GLB and OBJ formats.
• Implemented multi-view geometry, camera pose estimation, and structure-from-motion workflows for 3D asset generation.
• Reduced manual 3D editing effort by 40% through improved automated reconstruction.
• Enabled scalable 3D asset creation for e-commerce clients, lowering operational costs by 35%.

BASIS National ICT Award 2020 Champion · APICTA 2021 Finalist

• Built a salient-object-detection model reaching 96.23% accuracy using the HCE metric.
• Improved segmentation accuracy by 17% and processing speed by 30% compared with existing commercial tools.
• Optimized inference to process 257 MB images in 2.27 seconds using RSU blocks and advanced pooling.
• Scaled the platform to 5.7M+ paid images processed globally, with 8K-11K images processed per day for enterprise e-commerce clients.

• 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.

• Deployed SeamlessM4T for audio-to-text transcription in customer-service workflows.
• Integrated Llama 2 to extract contextual key phrases from transcribed conversations.
• Reduced manual transcription effort by 80%, enabling faster quality assurance and review.
• Built API workflows for transcription, phrase extraction, and downstream QA integration.

Production client: leading European luxury resale marketplace

• Designed and built KTM, a .NET and SQL Server platform that automates the full image-processing workflow from client FTP ingest to final delivery.
• Architected a 12-step production pipeline covering FTP upload, order creation, editor auto-assignment, AI background removal through Omnimage.ai, editor review, QC approval, and automated FTP delivery.
• Processed 4.5M+ images to date at approximately 200K images per month and 5K-8K images per day.
• Achieved an average batch turnaround of approximately 65 minutes through automated workflow routing and reduced manual handoffs.
• Built the order-management core to distribute files across active editors, collect outputs, and support client-facing QC approval.