— DeepTech AI Engineering · Sovereign AI · Malta EU
Less time searching. Faster decisions. Fewer manual handoffs. We build production AI for demanding workflows, sensitive data, and teams who need it to work. Start with a real problem. Prove the value. Then scale.
RetroGradient Labs is a member of the NVIDIA Inception program.
Accepted into NVIDIA's global startup program.
What could change
A clearer answer. An earlier warning. A smoother handoff. Start with the improvement your team needs.
Turn time spent searching into answers your team can check and use.
A policy question becomes a search through shared drives, old emails, and colleagues' memories.
Ask in plain language. Get a cited answer from the approved library, with access permissions intact.
Start with an approved document collection and the questions your team asks most often.
Turn existing camera feeds into useful signals about movement, activity, safety, vehicles, and physical spaces.
Useful events and patterns are buried across hours of video, noticed late, or reconstructed after the fact.
Interpret existing camera feeds in real time, surface relevant events and patterns, and route the result where it is useful.
Start with one high-value question, a defined set of feeds, and a measurable operational outcome.
Prepare the next step automatically, with people in charge of the decisions.
A request arrives by email. Someone rekeys it, checks attachments, and chases the next approval.
Extract the useful information, prepare the next step in your existing system, and flag exceptions for review.
Map a bounded process, check the data, and agree which decisions stay with a person.
Potential applications. We test the value with you in your environment.
Find your starting pointDon't know where AI fits? We find it. Data not ready? We fix it. Ready to deploy? We build it. The engineers who help define the solution are the people responsible for making it work.
Map the workflow, understand the friction, and check the data. Agree what a useful improvement would look like before choosing the technology.
Start with a bounded workflow and agreed success criteria. Compare the result with how your team works today, then decide whether to continue.
Integrate with your systems. Keep data, permissions, and human review under control, with deployment requirements agreed from the start.
Plan monitoring, maintenance, and ownership before launch. Improve the system as your workflows change, with a clear path for support.
Sovereignty and governance are built into every stage.
See how a pilot works →What we build
Find knowledge. Understand your space. Connect your systems. Choose the starting point that fits.
Cited answers from your policies, contracts, and team knowledge. Access rules stay in place; sensitive documents stay under your control.
Explore GRIMOIR →Currently in active engagement with Maltese government and private sector organisations.
RetroGradient Labs was founded by two engineers who have spent nearly a decade building production AI for businesses that needed it to actually work. We started this company because the gap between AI promises and AI products keeps widening, and the only way to close it is to build properly from the ground up.
Natalia is a Maltese AI engineer with eight years of experience building AI that actually ships, in factories, classrooms, and creative studios. She holds an MSc in Artificial Intelligence from the University of Malta, speaks at NVIDIA's flagship developer conference, and was recognised across three categories at Malta's Best Businesswoman Awards 2025. At RetroGradient Labs she leads strategy, client relationships, and product, with one rule: AI is only useful if it's trustworthy enough to put into production.
Adrian is a Maltese AI engineer with over half a decade of hands-on experience building and shipping AI systems at scale, from real-time camera analytics on edge devices to AI-powered tools and automated trading systems. He holds an MSc in Artificial Intelligence from the University of Malta and is a peer-reviewed AI researcher, co-first author of a 2024 paper in Springer's Neural Computing and Applications on explainable, adversarially robust computer vision. At RetroGradient Labs he leads the technical vision with a builder's mindset and a deep conviction that robust, production-grade AI is what separates great ideas from lasting products.
Not a blog. A body of work. Engineering depth, conference stages, and peer-reviewed research from the founders, in one place.
Benchmarking the NVIDIA A2, A10, A40, A100, L4, L40, and H100 across a real production video-inference pipeline (PeopleNet, FaceDetect, face-mask classification). Practical findings on tracker choice, interval values, and where decoders become the true bottleneck.
Read on Medium →Strategic takes on where AI hype meets engineering reality: LLM limitations, the gap between trustworthy AI and AI marketing, responsible deployment in regulated sectors, and what enterprises actually need before adopting generative systems.
Read on LinkedIn →A learning-free, hierarchical computer-vision approach to biometric identification on retina and palmprint images. Achieves perfect classification on the VARIA and RIDB retina datasets and 97.54% accuracy on the IITD palmprint dataset, while remaining robust to decision-based black-box adversarial attacks and to partial matching at 80% image visibility.
Read the paper →Before you start
What to expect, what to bring, and how we decide whether the work is worth doing.
Talk through your use case →No. Bring the workflow, the frustration, or the question you keep coming back to. We help identify where AI could be useful and whether your data and systems can support it. A clear problem is a better starting point than a chosen model.
GRIMOIR helps teams find and use internal knowledge. BEHOLDR turns camera feeds into operational signals. Custom engineering covers workflows and integrations that need a tailored approach. Tell us what you want to improve and we will help you find the right fit.
Yes. We start with a bounded discovery or pilot, agree the workflow and success criteria, and compare the result with how things work today. The decision at the end can be to scale, revise, or stop. A pilot is a way to make that decision with evidence.
How a pilot works →That depends on the workflow, data readiness, integrations, and deployment requirements. Discovery defines the work so we can agree scope, fees, deliverables, and timing before it starts. A production build is a separate decision after the pilot has been evaluated.
That is part of discovery. We assess what is available, what needs cleaning or organising, and which connections are practical. You do not need to replace everything to explore a useful workflow; we scope the preparation and integration work explicitly.
Yes. We support on-premises and private or sovereign EU-hosted deployments. We agree the data path, access rules, retention, and review requirements with your technical and governance teams before choosing the setup.
Explore deployment and governance →These use cases reflect workflows we have the engineering capability and prior delivery experience to implement. They are informed by systems and projects our team has worked on previously, as well as the operational patterns where we know AI can produce measurable value. Where RetroGradient client outcomes are available and we have permission to share them, we will publish those separately with the relevant metrics and evaluation context.
An engineer responds within 24 hours. We clarify the problem, discuss whether there is a fit, and suggest a concrete next step. You do not need a technical specification to start the conversation.
Leave with a clear next step.
Tell us what is slowing your team down, or choose a starting point below. No technical brief required. We respond within 24 hours.