Sports Image Search Platform · AI Solutions · Integration & Automation
WONDR Media
WONDR Media produces short-form NBA videos in the United States. We turned their request for a one-off image script into WONDR Image Search, a web app their whole editing team uses to find the right player photo, in the right jersey and season, with AI tagging, background removal and a searchable archive built in.
Built with
- Next.js
- SerpApi
- Google Gemini
- Python image processing
- Firebase Firestore
- Firebase Storage
- Dropbox API
- DigitalOcean App Platform
- GitLab
Expertise
- Search workflows for non-technical editors
- AI image tagging with Gemini
- AI background removal
- Background job processing
- API cost control
- Source ranking and filtering
- Admin tooling
- Handover to the client's own accounts
Verticals
- Sports media
- Video production
- Content studios
Every WONDR video needs a set of player photos: a portrait, a shooting shot, a defensive moment, a celebration. Editors were finding them by hand, running the same Google Images searches for every video, skipping watermarked results, removing duplicates and renaming files before they could start editing.
The brief asked for a Python script to automate that. In the first conversations we found what the script would have missed. A photo has to match its moment: a player traded last week must appear in his new team's jersey, and an archive piece about a past season needs photos from that season. And seven to ten editors would be running it, none of them developers.
We also caught a problem before any code was written. Google's Custom Search API, the obvious data source, had closed to new customers. We moved the build to SerpApi, which also gave us the size and aspect-ratio filters the brief needed.
We recommended a hosted web app instead of a script, and proved the approach before building it. A proof of concept ran on three cases the client picked: a recently traded player, a historical season and a lesser-known player. It showed exactly where raw search results fail, and we designed around each failure. Trusted sports sources rank first and carry a badge. Watermark-heavy stock agencies are blocked before anything downloads. Editors review every candidate, so a weak result never reaches an edit.
WONDR Image Search works the way an editor thinks. Type one player or several, add up to three modifiers, choose a team and a season from 1960-61 to today, check the live query preview and run it. Results arrive on a review screen with a filter for each player and modifier. Keep the good ones and download them one at a time or as a ZIP, already sorted into folders and named the way the edit expects.
We designed cost control into the product rather than leaving it to a monthly bill. Modifiers are capped so one search cannot multiply into dozens of paid API calls. Every downloaded image is stored with the search that found it, so the team's own library is searched first and a photo that has been used before costs nothing to find again.
Admins manage users, players, modifier groups and the trusted and blocked source lists from their own panel, without a developer. When long searches started hitting the time limits of the serverless host, we moved the app to DigitalOcean's App Platform and removed a separate Python service and job queue the design no longer needed. Jobs complete reliably and the running cost dropped. The code is in WONDR's GitLab and every service is registered in their name. They own the whole system.
WONDR's editors now start every video from one search screen instead of a browser full of Google Images tabs. Team, season and source quality are handled before an image reaches them, and photos arrive organised, named and ready for Premiere Pro.
The team adopted the tool in March 2026, and we turned their first weeks of feedback into improvements straight away: a clearer selection control, one-click single downloads and selective bulk downloads.
Each search adds to a photo library the team owns, on infrastructure in its own accounts and with source rules its admins control. The client summed up the engagement in a five-star review.
The AI layer
With the search platform in daily use, we added the AI layer that turns it into the team's own image library:
- Background removal. We tested three approaches: a paid background-removal API, a Gemini vision model, and open-source Python image libraries. The API was simple but too expensive at WONDR's volume, so we built a hybrid of the model and the Python libraries, running as background jobs so editors can send a whole batch of images and keep working.
- AI image tagging. Gemini tags each image with the player, team, action and setting. Editors search their own library by keyword instead of paying for a new web search, and the same pass flags watermarks and filters out images that are not useful, such as product or merchandise photos.
- Dropbox library import. WONDR's existing photo archive stays in Dropbox, and a searchable index of it lives in Firebase, tagged by the same pipeline. Years of past work now come up in the same search as new results.
- In-app research assistant. Editors can ask Gemini a question before they search, such as which team a player was on in a given season, so every search starts with the right context.
None of this needed a new system. Because every image had been stored with the search that found it from the first release, tags, archive images and AI processing extend the library the team already had.
Client feedback
What the client said
Vipin was a great communicator, went above and beyond on making sure all our needs were met. Highly recommend working with him!
Talk to an engineer
Like what you see?
Tell us about the system you need. We reply with questions and a suggested first step.