Houston, We Have a Hallucination
summary
A production conversational assistant for Ukraine’s Diia.AI was built using Python, LangChain and Kubernetes on Google Cloud calling Gemini, with a custom RAG pipeline and vector store to support retrieval. The system includes MCP integration fixes, on‑the‑fly PII anonymization, custom intent classifiers and a bespoke guardrail layer (including defenses against Ukrainian prompt injection), and scaled from one engineer’s PoC to a hybrid cloud/on‑prem GPU deployment producing ~83,000 messages and 952 income certificates.
AI-written summary · about
chapters (13)
- 0:00Introduction
- 2:31MinCyfra AI center
- 8:51Architecture drivers
- 13:48Diya AI reference
- 18:05Diya AI launch results
- 23:51Testing and security
- 27:00Data prep and RAG
- 30:28Business goals and assistant
- 36:17Technical architecture
- 43:26Data security and GDPR
- 45:32Commercial solutions comparison
- 48:38MCP communication details
- 50:03Model limits and load plan
related talks
45mAI Chats: Challenges in Business Integration
48mUnlocking LLM Potential: From Exploration to Integration
18mAI Platform Q&A
13mAI Q&A Session with Puneet Singh and Kapil Dhiman
40mPannel Discussion | From Hype To Value: Real World Applications Of Generative AI And LLMs
44mAI Foundations: Mitigate Risks and Boost SOC Efficiency
25mLangChain with Azure AI: Building LLM powered Applications
33mScalable LLM Inference on Kubernetes With NVIDIA NIMS, LangChain, Milvus and Flu...