GIDS news media, articles, insights and virtual events educate and illuminate its audiences so they can be fully prepared to deal with the new realities at work and in their professions.
Get the latest event updates, and insights from today's leading voices.
< session />
Thu, December 10Production AI SystemsInfrastructure, Platforms & Scale
Most AI agent tutorials focus on capabilities. Production systems fail for entirely different reasons.
Agents that perform well in development often encounter challenges when exposed to real users, long-lived sessions, concurrent workloads, streaming interactions, and operational requirements. Authentication behaves differently under load, observability breaks across protocol boundaries, memory accumulates over time, and resiliency mechanisms designed for traditional services can produce unexpected outcomes in LLM-powered systems.
In this session, we build a complete production-ready agent platform live, starting from an empty directory and progressing to a fully observable, authenticated, streaming application. Every component is assembled during the session, with no slides, prerecorded demonstrations, or hidden infrastructure.
As the architecture evolves, we will examine the production incidents that motivated each design decision. Attendees will see how concurrent streaming workloads can trigger OAuth refresh races, why distributed traces often disappear at Server-Sent Event (SSE) boundaries, how uncontrolled memory growth can affect response quality, and how traditional approaches to retries, caching, and circuit breakers behave in agent-based systems.
The implementation combines an MCP server, a LangGraph agent, a backend-for-frontend layer, OpenTelemetry tracing, checkpointing, configuration management, and streaming user interactions. By the end of the session, attendees will see a working system handling live requests while exposing a complete trace from browser interaction through tool execution and response generation.
The focus throughout is on operational realities rather than framework features. Every pattern is tied to a real production issue, the investigation that followed, and the architectural changes used to address it. Attendees will leave with a practical blueprint for building agent systems that remain observable, maintainable, and reliable after deployment.
What You Will Learn:
Who Should Attend:
< speaker_info />
Tuhin Sharma is Senior Principal Data Scientist at Redhat in the Data & AI team. Prior to that, he worked at Hypersonix as an AI architect and at IBM Watson as Data Scientist. He also co-founded and has been CEO of Binaize (backed by Techstars), a website conversion intelligence product for e-commerce SMBs. He received a master's degree from IIT Roorkee and a bachelor's degree from IIEST Shibpur in Computer Science. He loves to code and collaborate on open-source projects. He is one of the top 20 contributors of pandas. He has 4 research papers and 5 patents in the fields of AI and NLP. He is a reviewer of the IEEE MASS conference, Springer nature and Packt publication in the AI track. He writes deep learning articles for O'Reilly in collaboration with the AWS MXNET team. He is a regular speaker at prominent AI conferences like O'Reilly Strata & AI, PyCon, PyData, ODSC, GIDS, Devconf, Datahack Summit etc.
Soham is a Principal Engineer at Red Hat, where he builds agentic AI systems that survive the jump from demo to production. Based in Bangalore, he owns the full stack, from React front-ends through vLLM serving on Kubernetes, with reliability, observability, and security built in from day one.
Soham's production work, from enabling AI for multi-tenant data infrastructure to open-source agentic templates now running across Red Hat's Data & AI organization, reflects a rare cross-domain fluency across backend, security, and AI/ML. A published author and speaker, Soham cuts through the hype on what it actually takes to run AI at scale: closing the demo-to-production gap, taming cascading failures in multi-agent systems, and the engineering discipline that turns fragile prototypes into systems you can trust.
