Raphael Kalandadze is co-founder and chief technology officer of Wandero AI, which builds AI agents for travel businesses. His central engineering challenge is ensuring agents actually deliver correct itineraries, pricing, supplier coordination, and customer communications after they leave controlled demonstrations and encounter real customers.
From machine-learning research to travel automation
Kalandadze worked on computer vision, physiological signals, animal-behavior analysis, EEG decoding, and medical-image segmentation while studying information technology at Georgian Technical University. He co-authored research on multimodal emotion recognition combining video and physiological signals through an attention-based architecture.
As a team leader and assistant researcher with the Muskhelishvili Institute of Computational Mathematics, he led a 2024 workshop on Georgian-language AI, addressing data collection, augmentation, language adaptation, and instruction tuning. His public machine-learning projects include Georgian automatic speech recognition.
At Wandero, Kalandadze applies that background to agents that produce tangible business deliverables: itineraries, pricing spreadsheets, proposals, presentations, PDFs, and supplier confirmations. Consequential customer-facing actions retain human approval boundaries. Wandero’s engineering team rebuilt its production platform around more general-purpose agents after its specialized multi-agent architecture accumulated brittle handoffs and customer-specific exceptions.
Making agents reliable after launch
Production agent observability: A completed session can conceal serious failure: an agent might recover from an intermediate mistake by chance, select the wrong supplier, or miscalculate a customer’s price. Kalandadze evaluates execution trajectories and customer outcomes, not simply error rates or completion status.
Trace-to-pull-request remediation: Monitoring agents investigate production logs alongside the codebase, separate root causes from symptoms, and propose targeted fixes. An independent review agent critiques assumptions and edge cases before a human assesses the change.
Two operational feedback loops: Fast monitoring detects individual failures and urgent incidents; broader conversation analysis tracks session quality, costs, tool usage, recurring failure clusters, and longer-term system health.
Browser-based agent testing: Automated agents exercise actual product flows from the customer’s perspective, exposing interface and rendering problems invisible in backend logs. Kalandadze acknowledges the additional latency and token costs.
Wandero’s agent infrastructure combines isolated execution environments, scoped tools, approval workflows, contextual memory, and controlled credential access. Kalandadze’s AI Engineer World’s Fair presentation frames these safeguards and feedback loops as the missing layer after launch: the operational harness that turns deployed agents into systems their creators can understand, correct, and trust.
