AROBS Transilvania – custom software development company
Blog AI-Driven Precision in Syringe Filling for Pharma and Laboratory Applications
AI-Driven Precision in Syringe Filling for Pharma and Laboratory Applications
According to Data Bridge Market Research, the global medical automation market is valued at around USD 52 billion in 2024 and is projected to exceed USD 110 billion by 2032.
For pharmaceutical and laboratory manufacturers, innovations of this kind are not only about efficiency; they directly impact patient safety and regulatory compliance. AROBS recently collaborated on a project that applied AI to one of the most delicate processes in the pharmaceutical industry: syringe filling, where high precision and strict tolerances are essential.
Challenge
Pharmaceutical and healthcare operations run under extremely tight tolerances, where even slight deviations can lead to patient safety risks, costly product recalls, or regulatory non-compliance.
In this project, the client was looking to add a machine–learning–based supervision layer to the existing workflow. The objective was to provide real-time, complementary monitoring that could validate liquid volume with error margins under 1% (≈0.3 mL at 50 mL), using a standard laboratory setup (HD camera, workstation).
Gravimetric (weight-based) checks remain the validated industry standard, since image-only methods cannot account for liquid density. However, while accurate, gravimetric checks increase cost and complexity when scaled. The client therefore sought an additional safeguard – a “double-safety” AI layer – to visually supervise each fill, create a continuous audit trail, and enhance redundancy without replacing existing measurements.
Because syringes produced in manufacturing ultimately enter clinical use, the solution needed to be versatile, supporting both production-line quality assurance and clinical workflows.
The key stakeholders’ focus reflected this dual perspective: the Head of Manufacturing (concerned with waste and downtime), the Head of R&D & IT (focused on innovation), and the Quality & Compliance Officers (ensuring traceability). Any solution must integrate seamlessly into current operations, support GxP-aligned standards, and maintain errors within validated tolerances.
Solution
AROBS assembled a cross-functional team of AI engineers and life science compliance experts to design an AI-powered system that accurately estimates syringe liquid volumes, thereby complementing existing quality control methods.
Instead of building a complex multi-step imaging pipeline, we adopted a direct image-to-volume regression approach. In our controlled validation setup – with a fixed camera, consistent lighting, and syringe positioning in a jig – regression achieved the required accuracy while remaining efficient and robust. This made the solution easier to deploy in manufacturing and adaptable to healthcare environments, where clinicians need a simple, camera-based tool that works in real time without specialized hardware.
To ensure training and traceability, we developed an automated ground-truth capture process. We paired each image with a reference value from a calibrated syringe pump device, minimizing manual labeling and creating a audit-ready and compliant dataset. This ensured that every training sample had a trusted measurement attached, so the model learned from verifiable ground truth.
After evaluating several CNN architectures, we selected ConvNeXt-Tiny as the backbone, as it provided the best trade-off between accuracy and inference speed for our context. By framing the task as regression rather than object detection or OCR on gradation marks, we avoided unnecessary complexity and dependency on syringe markings.
The solution was engineered with practicality and scalability in mind. It runs in real time (~150 ms per prediction) on standard hardware, meaning the client does not need specialized GPUs or costly equipment on the factory floor.
Throughout development, we followed industry best practices for medical software, including versioned documentation, validation records, and audit trail support. The client’s Quality Assurance and Regulatory teams were able to evaluate and, if desired, approve the system for use in GxP-aligned environments.
Execution
Phase 1: Data Collection & Preprocessing
The first step involved establishing a controlled data capture environment using a calibrated, medical-grade syringe pump. The pump controlled the plunger position and documented liquid volume readings via SDK. Each camera snapshot was synchronized with the pump’s output: the capture software logged both elements together, saving the image and embedding the trusted volume value into the filename. This created a dataset where every image had a verifiable ground-truth reference, ensuring total traceability for training and audits.
To maximize learning efficiency, we applied preprocessing directly during capture. We defined a fixed Region of Interest (ROI) around the syringe barrel and plunger, trimming each high-resolution frame to the relevant area. It reduced background noise and highlighted only the critical visual cues of the liquid level. By feeding the model only for this focused ROI, we strengthened prediction reliability.
The resulting dataset included hundreds of images captured across multiple syringe positions and rotations, under consistent lighting, with and without visible gradation marks. Its balanced design equipped the model with diverse yet controlled samples, helping it generalize while maintaining precision.
Phase 2: Data Augmentation
To prepare the model for real-world variability, we applied targeted data augmentation to the captured dataset. Starting from a few hundred original images, we introduced controlled transformations that reflected the kinds of variations likely to occur in practice. These comprised tiny rotations to mimic syringe tilt, slight shifts in position or zoom, adjustments in brightness and contrast to emulate different lighting conditions, and minor noise or blur to reproduce camera imperfections.
In some cases, we also used color jitter to account for sensor or environment differences. Through these augmentations, the dataset was expanded to well over a thousand effective images. More importantly, the process ensured that the model would remain robust and reliable at inference time, even when encountering conditions slightly different from the controlled training environment.
Phase 3: Model Training & Validation
We trained the ConvNeXt-Tiny CNN regression model on the prepared dataset, with the syringe volume (in milliliters) as the target output. During the training phase, we monitored performance using metrics such as Mean Absolute Error (MAE), meaning that the model was learning to predict volume within the required precision range.
To track progress and detect issues prematurely, we analyzed loss curves and compared predicted values against actual ground-truth volumes, which helped us identify potential outliers or systematic biases. The model converged quickly thanks to the quality of the dataset and the focused preprocessing approach.
We iterated over hyperparameters to fine-tune accuracy, and kept the best-performing checkpoint along with a scaling factor to map outputs precisely to milliliter values. Internal validation tests confirmed that the model achieved error margins well within the target tolerance (≈0.3 mL at 50 mL, i.e. <1%), consistently meeting the project’s accuracy goals in our controlled test environment.
Phase 4: Deployment & Validation Tools
In the final stage, we built a lightweight validation tool to test trained models and demonstrate real-time performance. The application, developed in Python using Tkinter and OpenCV, permitted our team to load different model checkpoints, connect to a standard HD camera, and run inference on captured frames. At the click of a button, a snapshot was taken, the predefined ROI applied, and the predicted volume displayed on-screen. Each prediction could also be saved together with the input image for audit and compliance traceability.
This internal tool enabled systematic benchmarking of latency, accuracy, and robustness across multiple model versions. In our controlled tests, the system delivered predictions in ~150 ms per frame on a standard laptop, confirming real-time feasibility.
For the production environment, the AI component is designed to be integrated directly into the client’s controlled laboratory equipment and edge devices, providing continuous visual supervision as a complementary safety layer. This ensures that the solution can run reliably on-site without requiring specialized GPUs or significant workflow changes.
Result
The result is a scalable solution that strengthens how the client ensures quality in syringe filling. With AROBS’s AI-driven supervision layer, each syringe can be validated visually in real time, complementing standard gravimetric checks. This redundancy reduces the risk of undetected errors, reinforces patient safety, and lowers the likelihood of costly batch recalls or compliance penalties.
During controlled trials, the system consistently achieved error margins within the target tolerance (<1%, e.g. ~0.3 mL at 50 mL). Its stable performance gave the client confidence to move towards integration in operational environments. As a result, the solution is becoming a key element of the client’s QA process, adding both safety and traceability.
Operationally, the system helps reduce the manual inspection workload and reliance on expensive validation runs, as visual checks can be performed in real time by a camera and computer. This translates into lower labor costs, less material waste, and enhanced audit readiness. Because data capture and ground-truth logging were automated from the outset, the project also demonstrated how automated pipelines can accelerate machine learning adoption in the pharmaceutical industry.
Notably, the client gained a competitive advantage by adopting an AI-based, compliance-aligned approach. In an industry where trust and reliability are paramount, being able to show that every syringe is double-checked, by both standard measurement and real-time AI supervision, elevates their quality reputation.
Looking ahead, the modular nature of the solution means it can be extended to related challenges, such as identifying syringe types or detecting anomalies under less controlled conditions (e.g. varied backgrounds or mixed production lines).
Beyond manufacturing, the same approach can be adapted for clinical use: – Assisting nurses and technicians in hospital pharmacies during syringe preparation. – Supporting point-of-care syringe validation in oncology or infusion centers, where dosing accuracy is critical. – Providing an intuitive, camera-based interface deployable in hospital labs, pharmacies, or even mobile care units without requiring specialized equipment.
Broader Use Cases of This Approach
While this project focused on syringe filling accuracy, the same AI-driven supervision method has potential applicability across a range of healthcare and other high-precision domains where liquid volumes and dosing must be tightly controlled.
In the Medical Field, the strongest opportunities are for:
– IV Bag & Infusion Monitoring: Real-time verification of fluid levels, helping reduce nurse workload and prevent errors.
– Blood Collection & Transfusion: Volume checks for blood bags, contributing to safer transfusions.
– Diagnostic Labs & Sample Preparation: Validation of pipetted volumes in PCR, ELISA, or sequencing workflows, enhancing reproducibility and reducing human error.
– Oncology & Speciality Drug Dosing: visual supervision support for chemotherapy or biologic preparations, where tight dosing tolerances are critical.
– Pharmacy Compounding: Supporting hospital pharmacies in preparing syringes and vials with high precision.
– Dialysis & Critical Care: Monitoring fluid volumes in dialysis systems or ventilator circuits.
– Medical Device QA: For cartridges, catheters, and drug-delivery devices, sustaining alignment with FDA/EMA regulatory requirements.
Beyond Healthcare
The AI-powered supervision approach can be adapted for many industries where precision, compliance, and efficiency are essential, such as:
– Food & Beverage: Liquid filling checks in bottling lines to help reduce waste and improve consistency.
– Cosmetics & Personal Care: Supporting volume control for perfumes, serums, and creams, reducing material loss.
– Chemicals & Speciality Manufacturing – Providing an additional layer of precision in filling solvents or reagents, where safety and regulatory compliance are critical.
– Automotive & Industrial Fluids – Helping validate fills of lubricants, brake fluids, and coolants, reducing the risk of performance or safety issues.
– Consumer Packaged Goods (CPG) – Enabling fill-level checks in detergents, cleaners, and other household products, reducing the chance of customer complaints or recalls.
Conclusion
AROBS’s collaboration with the client demonstrates how a business outcome-focused use of AI in life sciences can deliver measurable value. Rather than innovating for its own sake, we addressed a critical quality challenge in syringe filling with a solution designed to fit the client’s operational reality and compliance requirements.
With sponsorship from both IT and Manufacturing leadership, the project has already shown returns: enhanced safety through double-checking, improved efficiency by reducing manual effort, and a forward-looking capability that differentiates the client in a competitive market.
This success also provides a foundation for future extensions – whether in related pharma processes or other high-precision industries – always with the same principle in mind: AI must be practical, auditable, and aligned with regulatory expectations.
Looking to achieve similar breakthroughs in your pharma or healthcare operations? AROBS’s pharma-tech experts are ready to support you in building AI-powered, compliance-aligned solutions that strengthen safety, efficiency, and competitiveness.
// Let us be the partner that helps your business adapt to change.
Leave us a message for a digital upgrade!
// our recent news
Read our recent blog posts
Case study
AI-Powered Log Analytics for Medical Pumps: Harnessing LLMs for safer operations and compliance
Read More »
Adina Sarbu
October 7, 2025