22% More Paid Conversions with AI-Powered Visual Production

Our client is a B2B SaaS platform that lets e-commerce sellers generate professional product imagery using AI, eliminating the need for costly photo studios and retouchers. Pynest joined a 15-person team to design and build the platform's microservices architecture, financial core and async generation pipeline from scratch. Within months of launch, the platform cut generation error rates from over 15% to under 2%, accelerated content delivery by 5× and drove a 22% lift in trial-to-paid conversion.

85%

Faster history loads

86%

Fewer system errors

22%

More paid conversions

Client Overview

Our client is an AI-driven media generation platform built for e-commerce sellers and marketplace vendors. It automates the entire visual production cycle, from model and location selection to final image output, eliminating the need for physical photo shoots. Users can configure generation parameters in real time, manage tasks through a queue-based pipeline and scale content creation via integrations with external AI services. The platform includes a full monetization layer: subscription plans, internal credit balances and payment gateway integrations, making it a complete end-to-end solution for producing marketplace-ready product visuals at scale.

CompanyB2B SaaS

RegionEurope

IndustryE-commerce, AI

CompanyB2B SaaS

RegionEurope

IndustryE-commerce, AI

Business Problem

Situation Before

Prior to the platform's full launch, creating visual content for sellers was entirely manual and dependent on external contractors, specifically photo studios and retouchers. Producing an image set for a single product card took 3 to 5 business days, with the cost of one quality image reaching $50-100. Scaling product catalogs to new regions was economically unviable at this cost and pace. Early product iterations ran on unstable scripts with no LLM orchestration, resulting in error rates above 15% and generation wait times exceeding 3 minutes. There was also no automated billing or subscription system, making Pay-as-you-go monetization impossible and limiting customer LTV.

Core Pain

Platform ReliabilityThe platform was unstable, which caused frequent failures and interruptions.

Revenue BlockWith no financial core, the client could not build any revenue model at all.

Resource WasteAround 25% of generations were wasted on defective source images that were never validated.

Risks

Up to 30% of potential revenue lost in the first weeks of a trend due to slow content production
Continued GPU resource waste from unvalidated low-quality source images (~25% defect rate)
No subscription or credit system meant no path to scalable, recurring revenue.
Infrastructure instability and missing queue management threatened platform reliability at any meaningful user volume

Constraints

Backend response time during generation status processing could not exceed 500 ms (strict UX latency SLA)
Legacy monolithic module could not be decommissioned until all users were migrated, backward compatibility required throughout
Compressed time-to-market: financial core and subscription system had to be ready before the e-commerce peak season
ComfyUI worker node API changes or library updates could interrupt active generation queues at any time.

Project Goals

Set up automated image production that scales without relying on photo studios.
Add a full monetization layer with subscriptions, credits and pay-as-you-go options to lift LTV and bring in recurring revenue.
Cut the lead time for launching new product collections and shipping AI workflow updates.
Catch defective source images before they enter the pipeline so GPU cycles are not wasted on unusable inputs.

Pre-validate source images to prevent defective inputs from consuming GPU resources.
Pre-validate source images to prevent defective inputs from consuming GPU resources.
Pre-validate source images to prevent defective inputs from consuming GPU resources.
Pre-validate source images to prevent defective inputs from consuming GPU resources.

Backend latency for generation status stayed under 500 ms.
System error rate dropped from over 15% toward under 2%.
Generation history load time fell well below the 3.5 second baseline.
Financial transactions ran at 100% accuracy, with no race conditions on coin balances.
New AI workflows could be deployed independently of the main release cycle.

Solution

The Microservices used HTTP sync calls for Auth, Billing and Orders as well as RabbitMQ asynchronous queues to decouple ComfyUI generation from the core backend. The legacy backend-ab was packaged in an Adapter microservice to preserve ML capabilities while migrating. Cleaning, reporting and statistics were scheduled using Kubernetes CronJobs. Backend-ab circuit provided GenAI engineers with an isolated environment to develop ML processes without affecting Production Release Cycles.

Backend-prod managed user flows in the Business Logic cluster, where ComfyUI workflows went through A/B testing and a pre-validation service that checked the metadata, resolution and masks before GPU execution. Financial Core had coins-service with DB-level locking, subscription-service, payment-service and order-service. Infrastructure included JWT/OAuth2 authentication for web and Telegram, S3 proxy for content delivery, CRM sync for marketing events and CronJobs.

Using AWS as its cloud service provider, the platform stored data in S3, used RDS for relational database storage and deployed applications on an EKS. All of the Platform's services were created with Docker, deployed using Kubernetes and had CronJob resources created to run periodic jobs. GitLab was used as the source code repository and CI/CD tool to automate build and deployment tasks. All background jobs were processed by a RabbitMQ message broker and a Celery task queue was used to manage them. This architecture provided a stable foundation and reproducible system for scaling purposes.

Unified JWT/OAuth2 authentication was implemented in both web and Telegram clients. DB-level locking prevented race conditions with financial transactions. Pre-validation service rejected invalid or low-quality images prior to GPU pipeline processing. Canary deployments restricted the blast radius of new releases; RabbitMQ had a Circuit Breaker to prevent cascading failures from GPU worker overloads. Key design decisions were managed using ADRs with mandatory code reviews requiring minimum 1 approval before any merge.

Architecture decisions

Technical initiatives

Security & Compliance

Infrastructure & DevOps

Key Steps

Month 1

Discovery & onboarding

Codebase audit of legacy scripts, requirements capture with client, ADR framework setup, Scrum process initialization, Jira configuration, architectural decision to split microservices into three functional clusters.

Month 2-3

Architecture design & PoC

Microservices topology finalized, financial cluster schema designed (coins, subscriptions, orders), ComfyUI async pipeline prototyped, DB-level locking strategy validated for balance operations, auth service (JWT/OAuth2 + Telegram) scaffolded.

Months 3-5

MVP build

Core backend-prod orchestration service delivered, coins-service and payment-service live with webhook processing, subscription and trial logic implemented, pre-validation service (resolution, metadata, mask checks) integrated before GPU submission, S3 proxy layer for generated content delivery, frontend dashboard and generation UI shipped.

Months 5-7

Staged rollout

Backend-ab isolated instance deployed for A/B testing new ComfyUI workflows without production risk, Kubernetes CronJobs live for data cleanup and stats recalculation, CRM integration (crm-actions) syncing registration/payment/churn events, error rate driven from 15% to under 2% through pipeline hardening and retry logic.

Months 7-9

Stabilization & scale

Generation throughput optimized 5×, history load time reduced 80%, order-service ledger live for analytics dashboards, full regression coverage by QA, CI/CD pipeline hardened on AWS, trial-to-paid conversion lift of 22% confirmed in analytics.

1 / 5

Month 1

Discovery & onboarding

Month 2-3

Architecture design & PoC

Months 3-5

MVP build

Months 5-7

Staged rollout

Months 7-9

Stabilization & scale

Results & Impact

85%Faster history load timeFrom 3.5 s to 500 ms

86%Lower system error rateFrom over 15% to under 2%

22%Higher trial-to-paid conversionWithin first 3 months post-launch

75%Faster AI workflow deliveryFrom 4 weeks to 1 week

25%Less GPU compute wasteFrom invalid inputs screened out

Business impactThe platform transformed the client's core business operation, replacing a slow, expensive, contractor-dependent content workflow with a scalable, automated AI production pipeline. The fintech cluster unlocked a Pay-as-you-go revenue model that had previously been architecturally impossible, directly contributing to a 22% improvement in trial conversion. With GPU waste reduced and error rates near zero, the unit economics of each generation improved substantially.

Operational improvementsThe cycle of producing content was reduced from 3 to 5 working days for each product card to near-instantaneous content generation. The error rate was reduced from above 15% to below 2%, thus reducing the load of handling the errors by the team. The GenAI team got complete deployment independence without any dependency on the core platform's release cycles. Manual data maintenance was eliminated completely using Kubernetes CronJobs, and the integration with AmoCRM automated marketing event tracking, including registrations, payments and churn, without engineering.

Project Snapshot

Team size & roles15 FTE: 1 TL, 1 BA, 1 PM, 4 BE, 3 FE, 2 AI, 2 QA, 1 DevOps

Engagement modelDedicated team

Duration9 months

Core tech stack

PostgreSQLAWS pythonfastapigitlab ciComfyUIcelerysqlalchemyasyncio

See Other Works

Secure LLM Gateway for a 63,000-Person Enterprise

Secure enterprise AI access with centralized governance, lower latency, content controls, and full token usage visibility

Bond Settlement Cut from 2 Days to 30 Minutes

End-to-end capital markets platform automating bond issuance, trading, settlement, and investor servicing

API Response Time Down 80% for a Bloomberg Rival

Financial analytics platform delivering real-time market data, news processing, and scalable backend services for fintech

Power Station Analytics: Algorithm Suite Delivered

Real-time equipment analytics platform for a power station with automated calculations, monitoring, and reliable fault detection

Show all