Aayush is currently unavailable
Aayush Garg
Verified Expert in Engineering
Deep Learning Engineer and Developer
Gurugram, Haryana, India
Toptal member since August 4, 2023
Bio
Aayush is a Senior ML Engineer with 7+ years of experience building production AI systems for Toptal and Fortune 500 clients. He takes generative AI and LLM/VLM technologies from prototype to production, from multi-agent systems and AI automation pipelines to LoRA fine-tuning and model deployment at scale. Aayush brings deep research experience and a PhD in high-performance geophysical computing.
Portfolio
Jiffy
Image Processing, Fine-tuning, OpenAI GPT-4 API, Claude Code, Claude...
RenderWolf AI, Inc.
Artificial Intelligence (AI), Python, Diffusion Models...
SixSense
Training, Convolutional Neural Networks (CNNs), Image Processing, Transformers...
Experience
- Image Generation - 7 years
- Python - 7 years
- Deep Learning - 7 years
- Linux - 7 years
- PyTorch - 7 years
- Software Development - 5 years
- Generative Pre-trained Transformers (GPT) - 3 years
- Machine Learning Operations (MLOps) - 3 years
Preferred Environment
Linux, Visual Studio Code (VS Code), Git, Python, MacOS, C
The most amazing...
...thing I built is a multi-agent image editing system that turns customer edit requests into real-time AI-generated edits with instant live preview.
Work Experience
Senior ML Engineer
2024 - 2026
Jiffy
- Developed Smart Redraw, a multi-agent AI image editing system for customer design edit requests with multiple agents like Analyzer, Artist, and Router Agents working in parallel to deliver real-time AI-generated edits with live preview.
- Built an image enhancement pipeline powering a $100+ million revenue business, combining AI upscaling, vectorization, classification, and background removal to transform 50,000+ daily raw customer uploads into print-ready artwork for Direct-to-Film (DTF) printing.
- Managed ML model deployments on Replicate and RunPod, and led the infrastructure migration from Replicate to RunPod, achieving significant GPU cost reduction through right-sized workloads.
- Optimized LLM costs by 30% via DSPy-based prompt optimization to migrate the vectorization routing layer from fine-tuned GPT-4o to GPT-5-mini and fine-tuned open-source VLMs as fallbacks for business continuity during proprietary model outages.
- Trained an in-house background removal model (BiRefNet-based) with significant edge quality improvements and faster inference than commercial alternatives.
- Built multiple AI automation tools including an AI image audit system for print-readiness defect detection, customer support automation for claims classification and a Zendesk customer ticket analysis pipeline delivering weekly executive reports.
Technologies: Image Processing, Fine-tuning, OpenAI GPT-4 API, Claude Code, Claude, Image Generation, Image Segmentation, NumPy, Machine Learning Operations (MLOps), Amazon Web Services (AWS), AWS Deployment, LLM Reasoning, Large Language Models (LLMs), LoRa, Artificial Intelligence (AI), Stable Diffusion, Diffusion-based AI Models, Chatbots, Custom APIs, FastAPI, Gemini, Google Cloud Platform (GCP), OpenAI, Large Language Model Operations (LLMOps), RAG Architecture, Retrieval-augmented Generation (RAG), LangChain, Data Pipelines, Agentic AI, AI Agents
Senior ML Engineer
2024 - 2025
RenderWolf AI, Inc.
- Deployed multi-model ComfyUI inference endpoints (SDXL, Flux, WAN 2.1, and Trellis) and a GPU-based LoRA training endpoint on Modal with auto-scaling and S3 integration.
- Ran extensive LoRA fine-tuning experiments on SDXL and Flux for character consistency and style transfer for game assets and created best practice guides.
- Built custom ComfyUI nodes (S3 integration, model loading, and image processing) for production endpoints.
- Created custom ComfyUI pipelines for multi-stage asset generation combining LoRA models, ControlNet conditioning, and image-to-image generation.
- Benchmarked and optimized inference for image (SDXL and Flux), video (WAN 2.1), and 3D (Trellis and Hunyuan3D) generation models, achieving significant speedups via PrunaAI caching and TeaCache.
Technologies: Artificial Intelligence (AI), Python, Diffusion Models, Diffusion-based AI Models, Stable Diffusion, Text to Image, Point Clouds, LoRa, Videos, Image Generation, Image Segmentation, 3D Image Processing, Model Deployment, Deployment, Prompt Engineering, Training, Machine Learning Operations (MLOps), Claude Code, Claude, Gemini, Fine-tuning
Senior ML Engineer
2024 - 2024
SixSense
- Benchmarked multiple CNN and transformer architectures for semiconductor chip defect classification, establishing a ConvNeXt-based architecture as the top performer.
- Conducted systematic ablation studies on the chosen architecture, tuning image size, label smoothing, and augmentation strategies, improving defect classification accuracy over baseline.
- Explored domain-specific pre-training on chip inspection data, demonstrating significant accuracy gains for downstream fine-tuning on curated defect datasets.
Technologies: Training, Convolutional Neural Networks (CNNs), Image Processing, Transformers, Generative Pre-trained Transformers (GPT), Machine Learning, Deep Learning, NumPy, PyTorch, SciPy, Data Science, Computer Vision, OpenCV, SVG, Fine-tuning
Software Development Engineer IV
2023 - 2024
FlixStock
- Led the development of fast, optimized LoRA and DreamBooth-based Stable Diffusion solutions for generating identities and products with varying aesthetics and poses.
- Spearheaded a team to develop an instant AI background generator that enables the creation of multiple professional-looking backgrounds for products and people.
- Developed a virtual try-on product to help customers visualize clothes, makeup, and accessories without putting them on physically.
- Deployed a custom stable diffusion image generation pipeline using ray-serve.
Technologies: Deep Learning, Hugging Face, PyTorch, NVIDIA TensorRT, Model Development, Model Deployment, Team Management, Leadership, Large Language Models (LLMs), Stable Diffusion, Computer Vision, Amazon Web Services (AWS), Google Cloud Platform (GCP), Text to Image, Product Design, ETL, LoRa, DreamBooth, Image Generation, Deep Neural Networks (DNNs), OpenCV, Object Recognition, ChatGPT, OpenAI GPT-3 API, Prompt Engineering, OpenAI GPT-4 API, Ray Serve, OpenAI, Fine-tuning
Generative AI Expert
2023 - 2023
Toptal Client
- Developed an optimized prompt engineering approach to generate high-quality product descriptions and metadata for home decor and furniture items, significantly enhancing search indexing and retrieval efficiency.
- Utilized advanced multimodal AI models to generate a comprehensive product metadata system, significantly enhancing search functionality.
- Architected an MVP hybrid search system by integrating OpenAI GPT models with Elasticsearch, enabling more accurate and context-aware results for the home decor and furniture inventory.
Technologies: Python, Machine Learning, Data Science, Artificial Intelligence (AI), Machine Learning Operations (MLOps), Generative Pre-trained Transformers (GPT), Amazon Web Services (AWS), OpenAI GPT-4 API, Modal, Prompt Engineering, Deployment, OpenAI, LangChain, FastAPI
Software Development Engineer III
2022 - 2023
Flixstock
- Devised and executed a state-of-the-art generative AI approach for creating realistic faces and identities.
- Pioneered a GANs-based generative AI solution to seamlessly generate human portraits in any desired pose.
- Developed an innovative image enhancement tool for eCommerce products.
- Created a cutting-edge image enhancement tool explicitly tailored for optimizing eCommerce product visuals.
Technologies: Deep Learning, PyTorch, Generative Adversarial Networks (GANs), Docker, Model Development, AWS Deployment, Team Management, Team Leadership, Product Design, Pandas, ETL, Image Processing, Stable Diffusion, Hugging Face, Artificial Intelligence (AI), Google Cloud Platform (GCP), Amazon Web Services (AWS), Machine Learning, Computer Vision, Diffusion Models, Large Language Models (LLMs), Text to Image, Deep Neural Networks (DNNs), OpenCV, Object Recognition, ChatGPT, OpenAI GPT-3 API, Prompt Engineering, OpenAI GPT-4 API, Ray Serve, OpenAI
Research Geophysicist
2019 - 2022
Shell
- Designed and implemented an innovative deep learning autoencoders-based approach for accelerated denoising of subsurface data.
- Engineered a real-time fault segmentation solution that significantly improved subsurface interpretation capabilities.
- Developed a cutting-edge deep learning (DL)-based Hessian operator to effectively mitigate crosstalk in large-scale multivariable optimization, specifically in solving wave equations.
- Conducted an industry-leading feasibility study on uncertainty quantification in subsurface realizations for carbon capture and storage, employing state-of-the-art DL-based approximate Monte Carlo techniques.
Technologies: C, Python, PyTorch, Physics, Geology, Image Retouching, Image Segmentation, Image Processing, Software Development, Stable Diffusion, Hugging Face, Artificial Intelligence (AI), Docker, Machine Learning, Computer Vision, Diffusion Models, Image Generation, Deep Neural Networks (DNNs)
PhD Researcher
2015 - 2019
TU Delft
- Developed an innovative deep learning (DL)-based super-resolution technique that successfully removes spatial aliasing from subsurface data, resulting in enhanced image quality and improved analysis accuracy.
- Researched reservoir-oriented joint migration inversion (JMI-res) technology, a groundbreaking solution that substantially reduces subsurface optimization costs while accurately estimating elastic parameters.
- Deployed the developed JMI-res technology in different energy companies' software infrastructures.
Technologies: C, Python, MATLAB, Optimization, Project Development, Artificial Intelligence (AI), Machine Learning, Computer Vision, Object Recognition
Experience
Building LLMs from Scratch
https://github.com/garg-aayush/building-from-scratchI built the complete LLM pipeline from first principles, covering pretraining, tokenization, inference optimization, supervised fine-tuning, and reinforcement learning.
• Character-level GPT
• GPT-2 reimplementation with RoPE
• BPE tokenizer from scratch
• LLM inference with KV cache and speculative decoding
• Supervised fine-tuning for reasoning and instruction following
• Expert iteration for self-improvement in math
• Group Relative Policy Optimization (GRPO)
• Character-level GPT
• GPT-2 reimplementation with RoPE
• BPE tokenizer from scratch
• LLM inference with KV cache and speculative decoding
• Supervised fine-tuning for reasoning and instruction following
• Expert iteration for self-improvement in math
• Group Relative Policy Optimization (GRPO)
cv-skills: Claude Code Plugin for Image Processing
https://github.com/garg-aayush/cv-skillsI built a Claude Code plugin that provides a comprehensive suite of classical image processing operations, enabling users to perform complex CV tasks by simply describing their needs in natural language. The plugin includes multiple specialized skills covering format conversion, SVG rendering, resizing, cropping, rotating, color adjustment, image filters, edge detection and segmentation, and compositing/blending.
ComfyUI SVG-to-Raster Node
https://github.com/garg-aayush/ComfyUI-Svg2RasterI developed a custom ComfyUI node that loads SVG files and rasterizes them into images for use in generative AI workflows. It was built using CairoSVG for high-fidelity SVG rendering and packaged as a drop-in extension for the ComfyUI ecosystem, along with example workflows.
Spatial Aliasing Removal Using Deep Learning Super-resolution
https://github.com/garg-aayush/spatial-alias-removalDeveloped a cutting-edge deep learning solution that effectively eliminates aliasing, including discretization noise, from subsurface data in real-time. This product received widespread acclaim and recognition within the energy industry. Additionally, as part of a collaboration with the University of Amsterdam (UvA), I took the initiative to make the starter code and data publicly available through an open-source platform.
Interactive Finance Real-time Stock Price Tracker Dashboard
https://github.com/garg-aayush/finance_dashboard_exampleMade an interactive dashboard to track stock price movement with features to generate simple moving averages and candlestick pattern graphs for data analysis. The dashboard can track six companies' stock prices in real time.
Deep Learning Templates for Efficient Experiment Tracking and Management
https://github.com/garg-aayush/pytorch-pl-hydra-templatesDeveloped various deep learning templates using the PyTorch Lightning framework and Hydra to facilitate efficient experiment tracking and project management. Additionally, created template scripts to enable comprehension and execution of training on distributed systems.
Model Parallelism for Deep Learning Neural Networks Architectures
https://github.com/garg-aayush/model-parallelismIn the context of training large models, particularly for 3D image segmentation and reconstruction challenges, the issue of out-of-memory arises when the model's size exceeds the capacity of a single GPU. This project addresses this challenge by demonstrating the implementation of model parallelism, specifically for intricate networks featuring skip and residual connections, leveraging a pipeline mechanism to achieve efficient computation.
Education
2015 - 2019
PhD in High-performance Computing for Imaging
Delft University of Technology - Delft, Netherlands
2010 - 2015
Master's Degree in Physics
Indian Institute of Technology (IIT) - Roorkee, India
Certifications
AUGUST 2025 - PRESENT
AI Evals for Engineers & PMs
Maven
JULY 2024 - PRESENT
Mastering LLMs For Developers & Data Scientists
Maven
Skills
Libraries/APIs
PyTorch, NumPy, SciPy, OpenCV, Ray Serve, Pandas, Custom APIs
Tools
ChatGPT, MATLAB, Git, AWS Deployment, Plotly, ComfyUI, Claude Code, Claude
Languages
Python, Bash, C, Python 3, C++
Platforms
Linux, Docker, Google Cloud Platform (GCP), Amazon Web Services (AWS), Visual Studio Code (VS Code), MacOS
Paradigms
ETL
Storage
Data Pipelines
Other
Machine Learning, Deep Learning, Image Generation, Image Processing, Stable Diffusion, Hugging Face, Artificial Intelligence (AI), Computer Vision, Diffusion Models, Text to Image, Deep Neural Networks (DNNs), Prompt Engineering, Software Development, Team Leadership, Machine Learning Operations (MLOps), Generative Pre-trained Transformers (GPT), Large Language Models (LLMs), Object Recognition, OpenAI GPT-3 API, OpenAI GPT-4 API, OpenAI, Physics, Geophysics, Algorithms, Simulations, Generative Adversarial Networks (GANs), Model Development, Team Management, Product Design, Geology, Image Retouching, Image Segmentation, Optimization, Project Development, Signal Analysis, Stock Market, NVIDIA TensorRT, Model Deployment, Leadership, LoRa, DreamBooth, Fine-tuning, Natural Language Processing (NLP), Data Science, Modal, Deployment, LangChain, MLflow, Recording, Videos, Diffusion-based AI Models, Point Clouds, Open-source LLMs, LLM Evaluations, 3D Image Processing, Training, Convolutional Neural Networks (CNNs), Transformers, SVG, LLM Reasoning, Reinforcement Learning, Chatbots, FastAPI, Gemini, Large Language Model Operations (LLMOps), RAG Architecture, Retrieval-augmented Generation (RAG), Agentic AI, AI Agents
Collaboration That Works
How to Work with Toptal
Toptal matches you directly with global industry experts from our network in hours—not weeks or months.
1
Share your needs
Discuss your requirements and refine your scope in a call with a Toptal domain expert.
2
Choose your talent
Get a short list of expertly matched talent within 24 hours to review, interview, and choose from.
3
Start your risk-free talent trial
Work with your chosen talent on a trial basis for up to two weeks. Pay only if you decide to hire them.
Top talent is in high demand.
Start hiring