How Swiggy builds reliable agentic data co-pilots at scale
summary
Production system walkthrough of how Swiggy turns plain-English questions into executable, auditable plans: agentic workflows that call validators and SQL engines, a semantic/ontology layer plus historical query embeddings to disambiguate columns, RAG/few‑shot caching that cuts LLM generations 60–70%, and human‑in‑the‑loop guardrails with observability and cost controls. Deployed pieces include dashboards, real‑time pod alerts, and automated daily summaries; user metrics show 500+ active users, ~3,000 queries, and localized deployments reporting 30–60% GMV lift where the co‑pilot is adopted.
AI-written summary · about
chapters (17)
- 0:00Introduction
- 1:02Beyond text-to-SQL co-pilots
- 3:28Reliable AI agent design
- 7:21Ontology layer for data
- 10:11Monitoring unknowns
- 11:59Swiggy’s co-pilot features
- 15:51Go-to-market strategy
- 17:57Onboarding with personas
- 21:48Training prompt engineers
- 24:13Continuous optimization
- 26:21Adoption and impact metrics
- 28:27Reliability and trust
- 31:13Adoption via champions
- 32:14Scaling co-pilots
- 33:37Q&A: Alignment and job security
- 37:13Q&A: Metadata challenges
- 41:29Q&A: ROI justification
related talks
41mCapacity and efficiency problems in real-time logistics - and how data at Swiggy solves this
25mFrom Analytics to Agentic Apps
48mMastering the AI Shift: Adaptive Strategies for Software Evolution
24mScalable GenAI in Production
58mFrom AI Sceptic to AI Sceptic Who Ships
32mFireside Q&A
35mBeyond the Hype and Hyperbole
25mBeyond the Hype: Real-World AI Strategies