stage/01
AI-generation detection
Classifies whether an image was synthesized or camera-captured, with a confidence score per image.
// case study — built at Applied AI
FormFlow
Multi-brand form administration with AI image-authenticity verification — so quality teams can trust the photos consumers send them.
my role
Product design → build → deployment, end to end
client
Quality-management division of a major Polish food & beverage group
stack
Next.js · Go · PostgreSQL · Kubernetes
programme
KPT ScaleUp Booster III (EU/FENG co-financed)
01 the problem
Complaint photos can no longer be trusted.
A consumer-goods organization runs dozens of brand websites, each with its own contact and complaint forms — historically, every change meant a developer ticket. Meanwhile, the photos consumers attach as evidence were processed with no verification at all.
Modern image generation and editing tools have made fabrication cheap: invented product defects, applied damage, altered labels. Quality teams had no way to tell an authentic photo from a doctored one. FormFlow fixes both problems at once — form management as a product, and image authenticity as a built-in pipeline.
02 verification
Three independent checks per attached image
stage/02
Manual-edit detection
Estimates the edited surface area and flags local manipulations — cloning, splicing, fills — visualized as a heatmap.
stage/03
Metadata verification
Validates EXIF consistency: capture date, device, location, and save signatures.
Every submission gets a per-image verdict — authentic · suspect · doctored — in a verification log the quality team can act on.
03 platform
Forms as a product, not a ticket queue
multi-brand workspace
independent forms, domains, and routing per brand — one operations view across all of them
drag-and-drop builder
field palette, validation, draft-to-publish flow with versioning history
topic-based routing
submissions land with the right team, automatically
embeddable forms
served on brand domains with allow-lists; verification runs invisibly after submit
unified inbox
cross-brand submissions with filtering, status badges, and verification reports
04 infrastructure
Boring infrastructure, on purpose.
A single Go backend, a Next.js frontend, PostgreSQL for storage and queueing, S3-compatible object storage. Shipped as one Helm chart that deploys the same way as SaaS or on-premise on Kubernetes. Predictable, auditable, and dull — exactly what a system handling legal-grade evidence should be.