About this session
Most multimodal agents look great in a demo and fall apart the moment they hit real enterprise data, users, and compliance requirements. Shikhar walks through the practical path from proof-of-concept to a production system that a business can actually depend on — the architecture, the evaluation loop, the guardrails, and the operational patterns that make the difference.
Expect a grounded, engineering-first view: how to wire vision, text, and structured tools into a single agent, how to keep it reliable as scope grows, and where teams most commonly get stuck between PoC and rollout.
Key topics
- Reference architecture for multimodal agents that combine vision, text, and tools
- Evaluation and observability that actually catch regressions before users do
- Guardrails, permissions, and human-in-the-loop patterns for enterprise environments
- Deployment, cost, and latency trade-offs when moving from PoC to production
- Common failure modes on the PoC → production path — and how mature teams avoid them
What you’ll take away
- A clear mental model for what "production-ready" means for a multimodal agent
- A checklist you can run your own PoC through before pitching it as a real system
- Concrete patterns for evaluation, guardrails, and operations you can apply next week
- A realistic view of where enterprise teams win — and where they burn quarters
Full session abstract to be announced. Reserve your seat now — every registrant gets the finalized agenda by email as soon as it's confirmed.
