From research to production: scaling SOTA search & reasoning agents
As part of World AI Week 2026, join us for an evening of technical deep-dives and conversations focused on building, scaling, and maintaining production AI for high-stakes industrial domains.
Foundational AI is moving fast, but bringing cutting-edge research into stable, complex industrial environments is a key challenge for enterprise applications.
We'll discuss practical architecture patterns, design insights, and real-world lessons from maintaining state-of-the-art systems in production. Bring your technical questions for the QnA and let's continue the conversations over drinks!
What we'll cover:
No traces, no problem: Post-training a sub-1B search agent for industrial fault analysis. Industrial root-cause analysis is a multi-hop reasoning problem where data includes conclusions, but no traces. With SFT dead on arrival and strict on-prem compute limits, we used RLVR and dense reward shaping to adapt a sub-1B search-and-reason model. We’ll cover the model hunt, our GRPO training stack, and how a sub-1B agent can outperform larger baselines while learning to search less. - Sankar Jaydas
Keeping a production agent harness open to new research. How we prompt models keeps changing: from completion prompts to tool-calling, to multi-agents, and now loops that rewrite the harness itself. We needed to keep up without rewriting our agent service every two weeks, we arrived at config-first harnesses long before it became fashionable. We’ll share our design principles and demo configuring our open-source harness to implement two recent self-optimizing loops: NVIDIA’s AVO and Prime Intellect’s Prime Agent. - Fernando Rejon Barrera
Who is it for: AI Engineers, Technical Leads, ML Systems Architects, Software Engineers, anyone building production-grade AI.
Looking forward to seeing you there!
Important: By registering for this World AI Week affiliated event, you consent to the sharing of your data with the event organiser, Inspired Minds Media Ltd, for the sole purpose of running the event.