AI & Engineering Executive building safe, scalable systems for high-trust environments
Kanishka is a Sydney-based AI and engineering executive with a unique track record of scaling machine learning across both hyper-growth tech and high-stakes healthcare. Sitting at the intersection of platform engineering, data, and AI, Kanishka specializes in architecting systems that operate reliably when trust and accuracy matter most. Currently, as the leader of engineering, data, and AI at Sonder, Kanishka is deploying human-centered AI systems that provide care professionals with strict operational guardrails and deeper context. Before this, Kanishka spent eight years driving massive AI and organizational scale at Rokt. For attendees looking to bridge the gap between AI prototypes and enterprise reality, Kanishka brings deep technical expertise and executive insight into building production-ready systems that actually scale, work seamlessly, and earn user trust.
In the high-stakes environment of mental health, medical, and safety support, direct-to-user AI chatbots and third-party APIs introduce unacceptable risks to patient safety and data privacy. At Sonder, we believe AI shouldn't talk to the vulnerable user; it should whisper to the expert operator and it must do so securely. In this session, we will unpack the architecture and engineering philosophy behind the Sonder Copilot, a human-in-the-loop (HITL) system powered entirely by sovereign, in-house AI models. Designed to supercharge operators with real-time summaries, automated case notes, and dynamic triage, we will explore our journey beyond basic prompt wrappers to build a secure, highly aligned AI assistant. Key technical takeaways will include: Sovereign AI & Human-in-the-Loop: Why hosting our own models is critical for data privacy, and how we integrate operator feedback directly from the call center UI to create a continuous data flywheel. Fine-Tuning vs. Prompt Engineering: Why we chose to fine-tune our own sovereign models to overcome latency, token limits, and strict tone requirements. Model Alignment via DPO: How we use Direct Preference Optimization to teach the model empathy and align its outputs with our best human operators without complex RLHF. Context Engineering (RAG): Automating triage by injecting relevant SOPs and safety protocols dynamically into the context window. Continuous Evaluation: Building a robust "AI Judge" to grade fine-tuned models on custom rubrics (Empathy, Accuracy, Safety) before production deployment. Join us to learn how to build defensible, sovereign AI architectures that secure sensitive data and elevate human experts rather than replacing them.
