Pixels and compression
Comparison of JPEG compression blocks between the damaged area and the rest of the image, sensor noise uniformity check, detection of splicing, cloning and sharpness mismatches.
Real project · Cosmos Wave agency, Amazon Verified Partner · WavePix module
Refund honest customers. Weed out fraud.
Every claim with photos is examined like a piece of evidence: forensic image authentication, and a packaging audit to tell a real breakage from a staged one.
The problem
A claim is often decided on gut feeling. Too much suspicion, and you lose an honest customer. Too much trust, and you refund edited photos, staged damage or AI-generated images. Either way, the real cause of the damage stays unexplained, and the returns keep repeating.
The approach
A claim isn't just a dispute to settle, it's information. Every analysis produces three results: a reasoned decision, a ready-to-send customer reply, and a recommendation so the problem doesn't happen again.
The solution provided
- Calculates a fraud suspicion score from 0 to 100, grouped into four tiers: none, slight, moderate, strong.
- Assesses the integrity of each photo on three axes: AI-generated image, editing or compositing, staged damage.
- Breaks down every customer statement and rates it: confirmed, unverifiable, or contradicted by the photos.
- Writes a complete internal diagnosis for the team.
- Prepares a courteous customer reply that never reveals the inconsistencies detected, so a fraudster can't adjust their evidence.
- Audits the packaging (corrugation, cushioning, sealing) and recommends low-cost corrections.
- Exports a complete PDF report, for record-keeping or a dispute with the carrier.
Forensics: authentication techniques
Retouching and AI images
Detection of overly smooth textures, distorted labels, barcodes and logos, anomalies on hands, and background transitions typical of generated images.
Water damage
Beaded droplets on the tape signal an outside spray: an internal leak soaks into the box, it doesn't bead on the tape. This clue strongly raises the score. Also checks for inside/outside contradictions and the direction of drips.
Physical consistency
The direction and hardness of the light, cast shadows and reflections must match between the product, the box and the floor.
Protecting honest customers
A blurry photo, taken at night or in low resolution: the score stays moderate so as not to penalize an honest customer. When in doubt, the tool recommends asking for more details, without accusing.
Real example
Claim for contents that were open and spilled in the box. Both of the customer's statements are confirmed by the photos. The images are judged authentic: a suspicion score of 5 out of 100. The packaging analysis reveals a lack of internal cushioning and suggests three corrections, two of them low or very low cost for a high impact.
Sensor noise: uniform
JPEG compression: consistent
What the data lets you decide
Profitability
Only refund valid claims, without penalizing honest customers.
Customer
Reply fast, with a reasoned and consistent response.
Risk
Spot edited photos, AI-generated images or staged damage before refunding.
Organization
Handle every claim with the same method, regardless of who receives it.
Management
Fix the packaging at the source to reduce future returns.
Every well-analyzed claim prevents the next one.
What this demonstrates
The same tool protects the margin and the customer relationship. The difference lies in the scoping: don't stop at the decision, trace it back to the cause.
Under the hood
Lovable, Supabase, analysis Edge Function, AI vision model, PDF export.
Are claims costing you a lot?
Your role. Your business problem. Your current workflow.
Working remotely across France, Belgium, Switzerland, Luxembourg, Quebec, North Africa and West Africa, in French, English and Arabic.
me@karimmesloub.com