About the Seminar
This seminar covers the architectural, operational and retrieval strategies required to deploy enterprise-grade AI systems in production.
Drawing on *three years of intensive, hands-on R&D building and evolving a highly challenging enterprise AI virtual assistant platform*, the seminar shares lessons learned from taking GenAI from early experimentation to real-world production at scale.
Across three sessions, we explore the evolution from simple LLM chatbots to scalable, multimodal assistant platforms. We cover the backend architecture and operational controls required for multi-model governance, prompt management and observability.
Finally, we examine the practical implementation and evolution of production-grade hybrid RAG pipelines.
Who Is This Seminar For?
Software Engineers, Backend Developers, Systems & Data Architects, Tech Leads, AI/ML Engineers, and Cloud/DevOps Leads who are designing, building, or scaling production AI applications.
Key Topics
Building a Production - Grade AI Virtual Assistant
- From chatbot to production-grade AI platform
- Core architecture: backend, orchestration, state, APIs and real-time communication
- Adding RAG and enterprise knowledge
- Integrating Speech-to-Text and Text-to-Speech
- Supporting avatars, streaming and multimodal interaction
- Architecture evolution, scaling challenges and lessons learned from production
Operating a Production AI Platform - Models, Prompts, Configuration and Lessons Learned
- Designing for multiple LLMs: model abstraction, routing, fallback and avoiding provider lock-in
- Managing prompts as production assets: Git, database storage, versioning and controlled deployment
- Dynamic prompt construction using placeholders, Jinja2 and reusable prompt components
- Administrating the AI backend: centralized configuration, feature flags, model settings, environment management and operational controls
- Observability and governance: logging, tracing, usage monitoring, token consumption, latency, cost and troubleshooting
- Production best practices learned in practice: safe configuration changes, Scaling, rollback, testing, resilience and operational maintainability
RAG in the Real World — What We Learned After Putting It into Production
- How RAG works and why production systems need it
- Ingestion, chunking, embeddings and vector databases
- Where pure vector search succeeds — and where it fails
- Hybrid search: combining semantic and lexical retrieval
- Metadata, permissions, ranking and context construction
- RAG evolution, evaluation and lessons learned from real-world usage
Schedule
Seminar Program
Building a Production – Grade AI Virtual Assistant
How do you evolve from a simple LLM-based chatbot into a scalable AI platform combining conversation, knowledge, speech, and real-time interaction?
Based on three years of hands-on experience designing and running the AI backend for Sidekick Platform, this session presents the architecture behind a modern Virtual Assistant and the key decisions required to operate it at scale.
Operating a Production AI Platform – Models, Prompts, Configuration and Lessons
Production AI systems quickly move beyond a single model and a few hardcoded prompts. Models change, providers evolve, prompts become business-critical assets, and the backend itself requires strong operational control, configuration, observability and administration.
This session presents practical architecture and best practices for managing models, prompts and the operational backend of a production AI platform, based on lessons learned from real-world use.
RAG in the Real World – What We Learned After Putting It into Production
RAG often looks simple in a diagram: create embeddings, search a vector database, and send the results to an LLM. In production, retrieval quality quickly becomes one of the hardest and most important parts of the AI architecture.
This session follows the practical evolution of RAG based on real production experience.