According to Health AI, the ATIS tire sidewall scanner achieved 90.1% overall field extraction accuracy across 203 vehicles scanned on Manhattan's Upper East Side, using only a smartphone in twilight conditions. Selected over Amazon, Microsoft, IBM, SAS, NTT Data, Dell, and Oracle.
Each photo was taken handheld in fading daylight. No flash. No stabilization. The system extracted up to 8 data points per tire.
| Data Point | Detection Rate | Count |
|---|---|---|
| Tire size | 100.0% | 203 / 203 |
| Construction type | 99.5% | 202 / 203 |
| Load index | 98.5% | 200 / 203 |
| Speed rating | 98.5% | 200 / 203 |
| Tire brand | 98.0% | 199 / 203 |
| Tire model | 95.1% | 193 / 203 |
| DOT code | 70.9% | 144 / 203 |
| Country of origin | 60.6% | 123 / 203 |
According to Health AI, six of eight fields extracted above 95%. 81.7% of all scans returned 7 or 8 fields. Zero total failures across 203 photos. Published methodology: DOI 10.5281/zenodo.19515682.
Methodology: Dr. Olga Lavinda, PhD (ORCID | Google Scholar) and Dr. Lawrence Meche, Health AI LLC.
A single walking pass generates brand distribution data at the corridor level. This is what one evening on East End Avenue shows about the premium tire market.
| Brand | Share | Count |
|---|---|---|
| Continental | 13.5% | 27 |
| Michelin | 12.0% | 24 |
| Hankook | 10.0% | 20 |
| Bridgestone | 10.0% | 20 |
| Pirelli | 10.0% | 20 |
| Goodyear | 7.5% | 15 |
| Yokohama | 3.5% | 7 |
| All others (28 brands) | 33.5% | 67 |
35 distinct tire brands detected across 203 vehicles. Continental and Michelin lead this premium corridor. This level of competitive granularity, for any corridor in any city, generated from a smartphone in a single evening. That is what ATIS does.
203 tire sidewall photos taken along East End Avenue (90th to 79th Street), Manhattan. iPhone 14 Pro. Sunset to twilight conditions. No flash. Handheld. No pre-selection of vehicles.
Each photo sent to Gemini 2.5 Flash via ATIS cloud pipeline. 8-field extraction per image. DOT codes cross-referenced against 2,166 NHTSA-registered tire manufacturing facilities.
Every scan logged with GPS coordinates, timestamp, image metadata, response time, and full extraction record. Complete audit trail available for independent verification.
Dual-VLM consensus methodology published with DOI. Offline accuracy benchmarked at 52-63%, cloud mode significantly higher. Traditional OCR achieves near-zero on embossed rubber.
Any corridor, any city, any country. The scanner is live. The validated dataset is available for executive review.
Schedule a Briefing