How We Are Building an AI-Powered Medical Device: The Vein Finder Story

The Vein Finder started with a simple observation: healthcare professionals struggle with vein access every day. An estimated 30 percent of first-attempt venipunctures fail in the general population, and for difficult patients — pediatrics, elderly, obese, dark-skinned, dehydrated, or chemotherapy patients — failure rates exceed 50 percent. Each failed attempt causes patient discomfort, wastes clinical time, increases infection risk from multiple punctures, and in severe cases leads to complications like hematoma, nerve damage, or delayed treatment.

We believed that combining infrared imaging with AI could solve this problem. Not just assist — actually solve it. The goal was a device that lets any clinician locate veins with the confidence of the most experienced phlebotomist, regardless of patient difficulty.

The Science Behind the Technology

The Vein Finder uses near-infrared light at 760 to 850 nanometer wavelength to image subcutaneous veins. At these wavelengths, deoxygenated hemoglobin in veins absorbs NIR light while surrounding tissue reflects it, creating contrast between veins and surrounding tissue. The physics is well understood — NIR vein visualization has been used clinically for over a decade.

What makes our approach different is the AI component. Existing vein finders on the market — AccuVein, VeinViewer — use NIR imaging without AI processing. They project raw or minimally processed NIR images onto the skin surface. The result works reasonably well for superficial veins on patients with light skin and low body fat, but performance degrades significantly for deeper veins, darker skin tones, and obese patients where subcutaneous fat scatters the NIR signal.

Our computer vision model, trained on over 8,000 clinical images across diverse patient demographics, compensates for these limitations. The model learns to identify vein patterns even when the raw signal is weak, noisy, or partially obscured by tissue scattering. It distinguishes veins from arteries based on depth and branching patterns. And it highlights the optimal puncture point — not just the vein location, but the specific point where the vein is straight, sufficiently superficial, and large enough for the required access.

Engineering Challenges: Where Software Meets Physics

Building a medical device is fundamentally different from building software. The reliability requirements are orders of magnitude higher. A software bug shows an error message. A medical device bug can cause patient harm. The engineering discipline required is closer to aerospace than to web development.

The optical system required months of engineering. Selecting the right NIR LEDs — wavelength, power, beam angle, thermal characteristics — determines the quality of the illumination field. The camera sensor must be sensitive in the NIR range while rejecting visible light. The optical filter must pass NIR while blocking ambient light that would wash out the vein image. And the projection system must accurately overlay the processed image onto the patient's skin with sub-millimeter registration.

We tested over thirty LED and sensor combinations before finding the configuration that provides consistent performance across skin types. The key insight was that a dual-wavelength approach — using both 760nm and 850nm illumination — provides better vein-tissue contrast than single-wavelength systems. The ratio between the two wavelengths varies with hemoglobin oxygenation, allowing our model to distinguish veins from arteries more reliably.

Thermal management is a non-obvious challenge. NIR LEDs generate heat, and medical devices have strict surface temperature limits to prevent patient burns. Our thermal design uses a combination of heat sinking, duty-cycle modulation for the LEDs, and thermal monitoring that automatically reduces power if surface temperature approaches the safety threshold.

The AI model runs on an embedded processor — an ARM-based SoC with a neural processing unit — that provides inference in under 30 milliseconds per frame. This enables real-time video overlay at 30 frames per second, giving clinicians a live, enhanced view of the subcutaneous veins as they position the device.

Skin Type Performance: The Equity Challenge

One of the most important engineering challenges — and one that existing vein finders handle poorly — is performance across skin types. NIR light penetrates and scatters differently depending on melanin concentration. Devices optimized for lighter skin types perform significantly worse on darker skin, creating a healthcare equity problem.

We addressed this by collecting training data across Fitzpatrick skin types I through VI, with deliberate oversampling of types IV, V, and VI. Our training dataset includes clinical images from Sri Lanka, the Middle East, East Africa, and South Asia — populations that are underrepresented in most medical device training data. The model learns skin-type-specific vein enhancement strategies, automatically adjusting processing based on detected skin characteristics.

In our validation testing, the device achieves first-attempt venipuncture success rates above 90 percent across all skin types — a significant improvement over existing devices that show 15 to 25 percent performance degradation on darker skin.

Regulatory Pathway: Navigating Multiple Markets

Medical devices must meet regulatory standards before they can be sold, and the requirements differ by market. We are pursuing regulatory approval in multiple markets simultaneously, which requires maintaining parallel documentation and compliance processes.

In the EU, the Medical Device Regulation requires CE marking through a conformity assessment process with a Notified Body. The classification is IIa — moderate risk — requiring a Quality Management System audit, technical documentation review, and clinical evaluation. In the US, the FDA classification for vein visualization devices is Class II, requiring a 510(k) premarket notification demonstrating substantial equivalence to predicate devices. In Australia and New Zealand, TGA registration follows the IMDRF framework.

The documentation requirements are extensive: design history file, risk management file following ISO 14971, biocompatibility testing per ISO 10993, electrical safety testing per IEC 60601, electromagnetic compatibility testing, software validation, clinical evaluation report, and instructions for use in multiple languages. We estimate the total regulatory investment at 300,000 to 500,000 USD across all target markets.

Clinical Testing and Results

We are currently in clinical testing with partner healthcare institutions in Sri Lanka and South Asia. The testing protocol compares first-attempt venipuncture success rates with and without the device, across patient demographics and difficulty levels.

Early results are encouraging. In a cohort of 200 patients across skin types III through VI, the device improved first-attempt success rate from 62 percent without the device to 93 percent with it. For the most difficult patients — elderly, dehydrated, or obese — the improvement was even more dramatic: from 38 percent to 87 percent. Average time to successful access decreased from 4.2 minutes to 1.8 minutes.

These results need to be validated in larger, multi-site studies before we can make clinical claims in our marketing. But the directional evidence is strong: the device meaningfully improves venous access outcomes, particularly for the patients who need it most.

What Is Next

The next milestones are completion of clinical studies by mid-2026, submission for regulatory approval in our initial markets, and establishment of manufacturing partnerships for production at scale. We are also developing the next generation of the device with additional capabilities — arterial detection, depth estimation, and integration with electronic health records.

The Vein Finder represents what Terra Labz is about: using engineering excellence to solve problems that actually matter. Not another SaaS dashboard. Not another social media app. A medical device that reduces patient suffering and improves clinical outcomes. If you are interested in our MedTech innovation pipeline — as a healthcare institution, partner, or investor — we would love to connect.