Team AI Research · Infobip
Research that carries a large share of the world's business communication.
AIR is the R&D team at Infobip. We work on messaging, voice and customer engagement, but focus on applied research that has an impact on live traffic at scale.
01 What we work on
Human-AI Collaboration
People trust AI communication systems unevenly. We study where that trust forms, and what it means for transparent, predictable design.
Trustworthy AI
Fairness, transparency and accountability in AI-driven communication, with evaluation and explainability built for real-world deployment.
Conversational AI
Most systems respond without resolving. We work on context, personalization and multimodal signals across text, voice and vision.
AI-powered Communication
Intelligent routing, adaptive orchestration and predictive network management that reduce latency and operational overhead, and recover from failures automatically.
Fraud Detection
Anomaly detection, behavioural modelling and prediction to catch fraud and malicious activity across communication channels, in real time and at scale.
Spam Filtering
Spam shifts faster than the models that catch it. Detection across text, images and voice, at the latency high traffic demands.
Voice AI
Speech enhancement, noise suppression, compression and unbiased voice processing, so voice quality holds up under real network conditions.
Generative Models
Domain-specific generative architectures for communication platforms, prioritising privacy, efficiency and operational control, including lightweight and privacy-preserving deployment.
More from the team — including posts and updates — also lives on research.infobip.com.
02 The team
Check out Ante’s blog — some of our ongoing research is updated there.Aug 2026Anchoring Bias in LLM-as-a-Judge Systems
03 How the team works
Industrial research fails in two predictable ways. It either drifts far enough from the product that nothing ships, or it collapses into feature work with a research label. The team is organised to avoid both: every area has a live question a peer would recognise as research, and a route to production where the answer gets tested against traffic.
That constraint shapes what counts as a result. An offline metric is a hypothesis. The finding is what happens when real users, real latency budgets and real distribution shift get a say — and often the interesting part is why the two disagreed.
We publish either way. A negative result that saves another team six months is worth more than a benchmark win nobody can reproduce, and writing it down is how a research team compounds instead of repeating itself.
04 Notes from the teamall updates →
Work with us
Joint agendas, pilots, theses
If you are working on something adjacent — a paper that needs an industrial testbed, a student thesis with real data behind it, or a joint agenda — that is the conversation worth having.
Get in touch