Dresma
Serving GenAI imagery to 50,000+ brands on fewer than half the GPUs
I architected the platform behind Dresma, a GenAI e-commerce content SaaS serving the US, UK and Europe, and led its entire tech org. The company has since grown to US$2M ARR on the same infrastructure.
- 12 to 5GPU servers for the same output, about 58% fewer
- 9M+images for 50,000+ brands, about 500,000 images a month
- < 1 minfirst images on the premium tier, against a 2-hour SLA
- US$2MARR the company has since reached on the same infrastructure
Challenge
Generating images with GenAI models takes GPUs, and GPU capacity was limited. DoMyShoot and the batch tools had to serve brands across three markets inside a 2-hour SLA, while the company also needed to reduce cost.
Approach
Multi-tenant GenAI pipelines
I architected cloud-native, multi-tenant pipelines behind DoMyShoot and the batch tools: Node.js on the backend, React on the frontend, Python for the models and Java for automation.
A tiered priority queue instead of more GPUs
Premium and standard queues ran across multi-cloud GPU capacity and cloud credits (AWS, GCP, Azure and private GPU partners). Premium users got first images in under a minute. Standard users got them in 10–15 minutes, well inside the 2-hour SLA.
In-house image models
We benchmarked open-source models and trained custom models on synthetic data.
- Segmentation
- SAM, U2-Net, BiRefNet
- Detection, smart cropping
- YOLO, Grounding DINO
- Category classification
- OpenAI CLIP
- Generation
- Stable Diffusion, SDXL
- Listing text
- Ollama self-hosted LLMs
Cost and organisation
Infrastructure streamlining, deployment automation and performance tuning cut hosting and processing costs by up to 30%. I also led the cost reduction and team resizing.
- Image jobsDoMyShoot and batch tools
- Priority queuePremium and standard tiers
- GPU capacityAWS, GCP, Azure, private GPU partners, cloud credits
- First imagesUnder 1 min premium; 10–15 min standard
- 2-hour SLA
- 120 min
- Standard tier
- 10–15 min
- Premium tier
- under 1 min
Results
- Scaled DoMyShoot and batch tools to 50,000+ brands and 9M+ images, about 500,000 images a month.
- Reduced GPU servers from 12 to 5 (about 58% fewer) for the same output.
- First images in under 1 minute on the premium tier and 10–15 minutes on the standard tier, against a 2-hour SLA.
- Cut hosting and processing costs by up to 30%.
- The platform still powers Dresma, which has since grown to US$2M ARR on the same infrastructure, after my exit.
- Company-reported Brand impact of up to 60% conversion uplift and 85% lower content costs. These are Dresma's figures, not ones I measured.
Stack
Public references
- Top Founders podcast, "How Dresma Hit $2M ARR With Usage-Based Pricing"
- The eCommerce Edge podcast, E471