Standard AI models fail on obscure medical equipment. Our custom vision model analyzed over 100,000 real-world images to identify manufacturers, models, and defects that generic AI simply can't see.
Snap a photo. Get instant device identification + defect analysis.
ChatGPT and Google Vision were never trained on medical device auction photos. They see a "hospital bed" or "medical monitor" - not the specific Hill-Rom P3200 Versacare or Philips IntelliVue MX800 that matters for service, parts, and compliance.
Useless for service techs who need exact manufacturer, model, and part compatibility information.
Actionable data: find service manuals, compatible parts, check recalls, estimate repair costs.
Building the world's most accurate medical equipment classifier required solving problems no one else has tackled. Here's how we turned 300,000 messy auction photos into a precision AI model.
Aggregated ~300,000 images from public online auction listings. The challenge: bad titles, missing brands, stock photos mixed with real equipment photos.
We didn't trust auction titles. Every image was sent to Gemini 3 with a specialized prompt: "Identify the Manufacturer, Model, and any VISIBLE DAMAGE (rust, cracks, discoloration)."
First attempt: Train strictly on damaged items. Filtered for rust, cracked screens, and dents. Problem: Dataset too small (~27K images). Model learned to find rust but forgot what normal equipment looked like.
Added 260K "clean" images. New problem: Model became lazy (90% clean data = always predicted "no defects"). Solution: Deep cleaning - synonym merging (zoll_medical to zoll), junk tag removal (stock_photo, new_product).
Balanced class imbalance (no_defects capped at 15K, kept 100% of defect images). Metadata fingerprinting to remove duplicate auction re-postings. Result: 82K to 43K images (-48%), dramatically increasing information density.
Smart balancing: Protect rare items (keep 100% if under 200 examples), sample common items down to 15K. Final result: 52,482 images, 329 classes, 100% validated labels.
Our custom model doesn't just identify devices - it provides actionable intelligence for service, procurement, and compliance workflows.
Recognizes 150+ medical device manufacturers including Hill-Rom, Stryker, GE Healthcare, Philips, Siemens, Medtronic, and dozens of obscure brands that generic AI has never seen.
Goes beyond brand recognition to identify specific model numbers and series variants. Distinguishes a P3200 from a P500, an MX800 from an MX700 - critical for parts compatibility and service procedures.
Automatically flags rust, cracked screens, dents, discoloration, worn casters, and other visible damage. Essential for auction buyers evaluating equipment condition and service techs documenting repairs.
Accurate identification unlocks powerful automations. One photo instantly links to FDA recalls, owner's manuals, work orders, recent sale prices, warranty status, and service history. The better the ID, the more actions become possible.
Snap photos during receiving to auto-populate inventory records
Identify unknown equipment on-site to find correct service docs
Evaluate condition and identify models before bidding
Document equipment conditions with AI-verified records
Our vision model is just the beginning. Here's what we're building next to make EzBot Vision even more powerful.
Training a Transformer to read price context ($50 vs $50,000) to improve identification accuracy and detect potential pricing anomalies in auctions.
Using image embeddings (vectors) to find duplicate photos even when auction titles are completely different - essential for market analysis and inventory tracking.
Human-in-the-loop system where the AI flags low-confidence images for expert review. Continuously improves accuracy with real-world feedback from biomed technicians.
Shrinking the model to run locally on iPads and phones. Real-time "blurry photo" alerts in the warehouse. Works offline in hospital basements with no WiFi.
EzBot Vision is currently in private beta. Schedule a quick demo to see how our custom AI identifies medical equipment that generic models can't.
See the vision model in action with your own equipment photos
We'll set up your account and train you on best practices
Upload photos from any device - web, mobile, or API integration
100% free for biomeds and hospitals. No credit card required.
Our custom model achieves 90%+ accuracy on manufacturer identification and 85%+ on specific model detection - far exceeding generic AI models which often can't identify medical equipment at all.
The model is trained on real-world auction photos - not studio shots. It handles poor lighting, angles, and partial views. If a human can identify it, EzBot Vision probably can too.
For completely new devices, the model will identify the closest match and flag uncertainty. Our active learning system means new equipment gets added to the model as users submit corrections.