Hareesh Menon
Hareesh is a technical manager with 12 years of experience in mobile app design and development. Away from work, he’s usually out cycling.
Hareesh is a technical manager with 12 years of experience in mobile app design and development. Away from work, he’s usually out cycling.
Most product teams treat a third-party SDK as a shortcut: license it, integrate it in days, move on. For a while, that holds. Then usage grows, the roadmap speeds up, and the shortcut starts setting the pace of the product.
That is the situation a leading Transport Management platform, an existing Fingent customer, found itself in. Its mobile app relied on a commercial off-the-shelf (COTS) scanning SDK so field staff and customers could capture invoices, receipts, delivery documents and financial paperwork. The scanner worked. But it was quietly constraining the roadmap, inflating costs and adding risk in places nobody was tracking.
The client asked Fingent for a clear outcome: remove the dependency on the legacy scanner. What follows is a useful demonstration for any business weighing build vs. buy for a core capability.
The client was paying approximately $5,000 a year, billed on a volume basis, so it could only grow with usage. A five-year review of support history showed the real cost sat well beyond that.
The scanner capped each license key at 200 seats, with three keys per subscription period. Every fresh install consumed a seat, whether it came from a new user, a new build, or an ordinary reinstall after a device swap.
With no live license dashboard, only periodic emails, teams often found out about depletion after users had already begun submitting uncompressed JPEGs instead of the PDFs the workflow required. Past the cap, the camera preview returned blurred, unusable images, and recovery meant an unplanned emergency build.
The PDF conversion module occasionally failed on its own while capture and cropping kept working. Resolving it meant long coordination across several third-party support teams.
Because builds consumed licenses, the team could safely plan only three releases a year. Mandatory SDK compatibility updates, needed whenever the operating system moved ahead of the vendor, regularly took one of those slots, leaving fewer for features and fixes. In effect, a license pool was dictating the product calendar.
The client had 121 active users against a 200-seat key, which looks safe. But users reinstall, change devices and move in and out of the company, and each event uses another seat. With 20% annual user growth and about 30% of the base consuming an additional license each year, effective consumption reached roughly 157 in Year 1 and 189 in Year 2, then breached the cap in Year 3, about 14 months sooner than headcount growth alone would suggest.
Fingent translated the documented incidents into an annual estimate:
Further, this total cost still omits a rather harder-to-calculate damage. For a new user, or a returning user in a critical season, a broken scanner or a forced downgrade to a low-quality JPEG can erode the trust in the app.
Two more pressures sharpened the case. The client's business is seasonal, so sign-ups surge in December, January, June and July, exactly when the seat ceiling was most likely to break. And the client wanted to move away from a closed-source, consumer-grade SDK inside an app that handles sensitive business documents.
That concern is not abstract: in 2019, researchers found a hidden malicious component in a widely used commercial scanning app with more than 100 million downloads, and a national government later restricted a group of apps that included it. The lesson is general. When you embed a closed SDK, you cannot audit it, patch it or see how its vendor handles security.
Building is not the default answer. Before committing, Fingent assessed what the market already offered against a defined set of requirements:
Five alternatives were assessed one by one:
Removed per-seat licensing but kept the scanning UI and detection algorithm entirely in Google's hands. The client wanted no Google dependency in the document pipeline, and replacing one vendor with another would not meet the sovereignty goal.
Handled edge detection and cropping but had no native PDF conversion or rotation, limited UI flexibility and no active community.
Lacked PDF generation, rotation and support for modern frameworks such as CameraX and Kotlin, and its codebase was abandoned.
Had no PDF export or manual rotation, weak contrast tuning, hardcoded screens and an unmaintained repository.
Had no built-in PDF creation, rotation or contrast filters, and needed heavy custom glue code for a usable interface.
The pattern was consistent. Each option removed one constraint of the legacy scanner while leaving core needs unmet, above all PDF export, rotation, active maintenance and full ownership of the UI and detection logic. Whichever library was chosen as a starting point, Fingent would have had to close and maintain those gaps anyway. That is what tipped the decision toward a purpose-built solution.
The market context supported the call. Per-seat and volume-based licensing is becoming less common among newer commercial scanning SDKs. Scanbot SDK, for instance, is marketed on a flat annual fee that does not vary with users, devices or scan volume. Industry data on comparable mobile scanning SDKs also points to a typical three-year licensing cost of $15,000 to $45,000 before any hidden operational costs.
ScanCore is a proprietary document scanning engine designed to match the scan quality and experience the client's users already expected, without the constraints above. Its core idea is dual-detector fusion.
On every capture, a pretrained ONNX corner-detection model and a classical OpenCV contour-detection pipeline (Canny edge detection, Otsu thresholding and adaptive low-light handling) run in parallel. A confidence-weighted decision matrix, developed in-house by Fingent engineers, reconciles the two.
When both detectors agree closely, their results are averaged. When they diverge, the matrix picks the higher-confidence result, favoring OpenCV in low light and ONNX for larger documents. A validity check rejects any candidate with implausible corner angles before the image moves on to perspective correction, enhancement and export.
Around that engine, ScanCore delivers:
Automatic perspective correction and cropping, with an optional manual trapezoid-correction UI for edge cases
An enhancement pipeline with sharpening, CLAHE and multiple binarization algorithms (Sauvola, Wolf, NICK and Otsu), tuned to match the legacy scanner's output
Multi-page capture compiled into a single size-capped document or PDF
A custom scanning UI built and maintained by Fingent
No per-install licensing, no key management and no external processing pipeline
In internal testing on 1,000 document captures under typical indoor lighting, 94.6% were correctly bounded on the first attempt, 5.8% needed a retake (mostly because of low light or extreme angles), and the average capture-to-result time was 650 milliseconds on a mid-range Android device.
An honest build-vs-buy story needs honest numbers. Fingent audited the ScanCore codebase, roughly 11,000 lines across 34 source files, and estimated that one Senior Developer working full-time can build it in about 2.5 months. The estimate assumes the pretrained ONNX model is reused rather than developed from scratch.
One Senior Developer, ~11 person-weeks, or about 2.5 months of focused work.
Approximately $14,300 at a blended $1,300 per week.
Roughly 40 hours a year, approximately $1,300*
*Covering app store policy changes, upgrades to core libraries such as OpenCV and the ONNX Runtime, and framework updates such as CameraX
The build figure is an engineering estimate rather than a billed amount, and the maintenance figure is approximate.
Using the figures above, the picture looks like this:
| Measure | Calculation | Result |
|---|---|---|
| Legacy fully loaded annual cost | Calculation $5,000 + $6,300 + $7,700 | Result About $19,000 |
| ScanCore annual maintenance | Calculation 40 hours at the blended rate | Result About $1,300 |
| Net annual saving | Calculation $19,000 − $1,300 | Result About $17,700 |
| Approximate payback | Calculation $14,300 ÷ $17,700 | Result About 0.8 years (roughly 10 months) |
Two caveats. First, the payback is a conservative, flat view: the legacy cost is volume-based and grows with the user base, so in practice the break-even point arrives sooner.
Second, it depends on counting incident costs, not only the subscription. On the subscription fee alone ($5,000 less $1,300), the payback would stretch to just under four years. The case for building rests on the full cost of ownership, which is why that cost needs to be measured first.
The recurring $5,000 license is gone, replaced by roughly $1,300 of annual maintenance. The client also avoids the incident-related engineering effort and at-risk revenue, bringing total avoided cost to about $19,000 a year against that maintenance figure.
With no license pool to protect, the team moved from three rationed releases a year to a monthly cadence, a four-fold increase, with every release slot free for product work.
Sign-up surges in the busiest months no longer collide with a seat ceiling, so new users meet a working scanner at first impression instead of a broken one.
Document image processing stays inside the client's own application and infrastructure boundary, with no dependency on an outside scanning vendor's pipeline. For users scanning financial and business documents, that supports a stronger trust proposition.
Talk to our experts about reducing your SDK dependency and having a custom-built scanning tool bespoke to your needs.
ScanCore's footprint is larger than lightweight options such as ML Kit. Dynamic Play Feature Delivery, which loads the scanning module on demand instead of bundling it into the base install, has largely offset this.
Bug fixes and upgrades for OpenCV, the ONNX Runtime and CameraX sit with your engineering team rather than an outside vendor.
A COTS SDK can be integrated in days or weeks. A build of this kind takes months, so cost and agility benefits accrue over time rather than at once.
Performance on low-end and older Android devices needs ongoing tuning. About 3.5 person-weeks of the build went into cross-device and lighting testing, and the work continues as new devices arrive.
Building makes sense when the capability is central to your product, when licensing or vendor terms limit growth, and when you need control over privacy, UX and release cadence. Buying remains the better choice when speed to launch matters most, the capability is peripheral, or you lack the engineering depth to maintain it.
Before renewing your next SDK license, ask:
Fingent designs, builds and maintains custom software for businesses across transportation and logistics, healthcare, education and financial services. Our engineering teams work as an extension of our clients' product organizations, from mobile and web applications to specialized systems built on computer vision, data engineering and cloud infrastructure.
The ScanCore engagement shows how we approach the work: understand the underlying business constraint, evaluate what the market offers, and build when building is the right long-term decision. ScanCore is also a reusable Fingent accelerator, which means organizations facing a similar dependency can start from a proven engine instead of from zero.
Talk to our experts about reducing your SDK dependency and having a custom-built scanning tool bespoke to your needs.
ScanCore is Fingent's in-house document scanning engine. It combines a pretrained ONNX corner-detection model with an OpenCV contour-detection pipeline, fused through a confidence-weighted decision matrix, to deliver edge detection, perspective correction, enhancement and PDF export.
Consider building when the capability is core to your product, licensing limits constrain growth or release speed, you have data sovereignty or privacy requirements, and the full cost of ownership over several years favors an in-house solution.
Fingent’s audit estimates about 2.5 months, or roughly 11 person-weeks, for one Senior Developer, assuming a pretrained corner-detection model is reused.
Mainly app store policy compliance, upgrades to core libraries such as OpenCV and the ONNX Runtime, and framework updates such as CameraX. For ScanCore this is approximately 40 hours a year.
Yes. ScanCore is a reusable Fingent accelerator that can be adapted to other products and requirements. Reach out to the team for a demo.
Years of Partnership:
Extended Team Size:
Years of Partnership:
Extended Team Size:
Years of Partnership:
Extended Team Size:
Years of Partnership:
Extended Team Size:
Years of Partnership:
Extended Team Size:
Years of Partnership:
Extended Team Size:
Sample Blueprint
Sample Blueprint
Contact Us