How long did it take to go from audit to a working computer vision app?

The full engagement ran four months, starting with a return-data and support-ticket audit to identify the highest-error part categories, then moving through model training, app development, and a staged rollout. Because the rollout began with the worst-offending categories (brakes, sensors, connectors) rather than the full catalogue at once, the client saw measurable accuracy gains well before the four-month mark.

What happens when the AI isn't confident about a match?

The model is integrated with the live catalogue index, so whenever confidence falls below a defined threshold, the app shows the top match alongside a ranked list of visually similar alternatives instead of forcing a single guess. This keeps low-confidence cases from silently becoming wrong-part returns, and every one of those edge cases feeds the retraining feedback loop.

Does this only work for automotive parts, or could it apply to other product catalogues?

The approach is built around any catalogue where visually similar SKUs cause costly misidentification — brackets, connectors, and fittings in this case — so it generalizes well to other technical or industrial parts marketplaces with the same problem. The core pattern (photo-first UX, confidence-scored matching, feedback-driven retraining) isn't automotive-specific.

What was the biggest risk in this project, and how was it managed?

The biggest risk was training a model on clean vendor catalogue photos that would then fail on messy real-world customer snapshots taken in garages and driveways. SystimaNX addressed this upfront by deliberately augmenting the training data with varied lighting, angles, and backgrounds, then validating accuracy on the highest-error categories first before expanding to the full catalogue.

Back to case studies

Automotive & Retail TechnologyAI-Powered SoftwareSaaS Application Development

AI Computer Vision Parts Marketplace

Eliminating manual part identification errors with a mobile-first marketplace powered by computer vision, replacing guesswork-driven text search with an instant photograph-to-SKU purchase flow.

Drastically reduced return rates via AI-driven part identification

Client

Confidential — Automotive retail platform

Industry

Automotive & Retail Technology

Timeline

4 months

Technologies

8+ tools

The Challenge

!The marketplace's manual part identification process had a high error rate, with customers routinely selecting the wrong SKU when trying to match a worn, corroded, or unlabelled component to a listing. Each misidentification triggered a full return-and-refund cycle, and the cumulative cost of restocking, re-shipping, and processing refunds was steadily eroding margins on already low-margin consumables.

!Customers were expected to identify their own parts by name or part number within a catalogue spanning thousands of SKUs across dozens of vehicle makes and model years. Many shoppers — particularly DIY buyers and first-time car owners — simply did not know the terminology, and support tickets asking staff to identify a part from a blurry photo became a recurring, unscalable burden on the customer service team.

!The existing text-based search engine was built for buyers who already knew what they needed, not for buyers trying to diagnose an unfamiliar part. Fuzzy matching and autocomplete helped marginally, but they could not resolve visually similar parts that differed only in bracket shape, connector type, or mounting pattern — precisely the differences that caused the most returns.

!Return processing was disproportionately expensive relative to order value. Automotive parts are often bulky, require special packaging, and can be damaged in transit twice — once on the original shipment and again on the return — so a single wrong-part order could wipe out the margin on several correct orders.

!Despite mobile devices generating the majority of site traffic, the purchasing experience was still optimized for desktop, with a dense multi-filter search UI that was cumbersome to operate on a phone screen and assumed users had the patience to browse rather than snap a photo and buy.

!There was no mechanism to learn from past identification mistakes. Customer service logs of misidentified parts sat in ticketing software, disconnected from the product catalogue and search index, so the same categories of errors recurred month after month with no systematic improvement.

!Vendor-supplied catalogue images were inconsistent in angle, lighting, and background, making any future computer vision effort harder — a naive model trained on this imagery risked overfitting to studio photos rather than generalizing to real customer snapshots taken in garages, driveways, and under car hoods.

!Leadership needed measurable proof that an AI-driven fix would move the needle before committing further engineering budget, but there was no existing instrumentation to track identification accuracy, return reasons, or mobile conversion by part category.

Our Solution

SystimaNX began by auditing return data and support tickets to quantify which part categories drove the highest error and return rates, prioritizing the computer vision effort on brake components, sensors, and electrical connectors where visual ambiguity was most costly.

We built a mobile-first native app for iOS and Android with a camera-based identification flow front and center on the home screen, deliberately relegating text search to a secondary path so the primary purchase journey started with a photograph rather than a query.

A computer vision model was trained on a curated, augmented dataset combining vendor catalogue images with real-world customer photos, deliberately introducing variation in lighting, angle, and background so the model would generalize beyond clean studio conditions.

We designed a deliberately simple three-step UX: photograph the part, let the AI analyze it in under two seconds, and present the exact SKU match with a one-tap add-to-cart — removing every unnecessary screen between intent and purchase.

The computer vision model was integrated directly with the real-time catalogue search index so that, alongside the top match, the app surfaced confidence scores and a ranked list of visually similar alternatives whenever certainty fell below a defined threshold.

A structured feedback loop was implemented so every mismatch flagged by a customer or support agent was logged, reviewed by the team, and folded into the next model retraining cycle, turning what had been an unresolved recurring problem into a continuously shrinking one.

We stood up an analytics dashboard giving product and support teams live visibility into identification accuracy, return rate by part category, and mobile conversion, replacing anecdote-driven decisions with a data feed leadership could act on weekly.

Finally, we ran a staged rollout — starting with the highest-error part categories identified in the initial audit — so the team could validate real-world accuracy and tune confidence thresholds before extending computer vision identification across the full catalogue.

Measurable Impact

Part identification accuracy

94%+

The computer vision model outperformed manual search accuracy across every tested part category, including visually ambiguous components like sensors and brackets.

Return rate reduction

35-45% reduction

Correct first-time identification cut wrong-part returns by roughly a third to nearly half in the highest-error categories, directly protecting margin on high-volume consumables.

Mobile conversion

3x lift

The camera-first mobile experience tripled conversion compared with the previous text-search flow.

Support ticket volume

40% fewer tickets

Misidentification-related tickets dropped by roughly 40%, as fewer wrong parts meant fewer 'is this the right part' escalations reaching customer service.

Model improvement

+2-3 points per retraining cycle

Each production retraining cycle, fed by the feedback loop, lifted model precision by roughly two to three percentage points on previously flagged part categories.

Time to purchase

Under 15 seconds

The three-step photograph-to-checkout flow cut the average path to purchase to under 15 seconds, down from several minutes of filter-based text search.

“Returns were eating our margins, and every wrong-part shipment cost us twice — once to send it, once to take it back. SystimaNX built something our customers love using, and the returns problem is practically solved. The accuracy numbers alone justified the investment within the first quarter.”
H
Head of Product
Automotive Marketplace (NDA)

Technology stack

React NativePythonTensorFlowAWS RekognitionFastAPIPostgreSQLS3Node.js