Talk to Your Data: Production-Grade AI Agents for Analytics & Finance
We had an incredible evening at Amsterdam’s premium Cinema The Pulse for the latest PyData Amsterdam meetup, hosted by Manychat!
We moved way past basic pipelines into the reality of shipping production-grade AI. If you are building with LLMs, this was a massive reality check on turning experimental text-to-SQL toys into bulletproof enterprise tools. How they came alive on the big screen?
🎬 Talk 1: Answers You Can Question: Building a Trustworthy Self-Service Analytics Agent
Alex L., Senior AI / Data Engineer at Manychat, pulled back the curtain on how his team designed an internal analytics agent on top of Claude Code. By deploying an open architecture that lets non-technical employees use a /ask command, they have successfully bypassed the traditional analyst queue to query metrics directly in plain English. Sitting in the audience, watching their engineering choices laid out on the big screen made it clear that building a trustworthy self-service agent relies on these fundamental strategies:
▪️ Two layers of context; a semantic layer and the brain; the first layer contains meta data context, and the second smart guard rails.
▪️ To prevent quiet breaks - re-check on a daily basis; routing, view citation and using numbers within tolerance.
▪️ Unsolicited advice; start deterministic, the repo is all it sees, and progressive disclosure. For the latter: when everything is important, nothing is. So yeah, from the start; less is more for good answer quality 😎
🎬 Talk 2: From Chat to Insight: Reliable AI Agents for Financial Data
Niels Neerhoff, Software Engineer at Palm, shifted the focus to the high-stakes world of treasury management and fintech. He broke down the core engineering challenges of applying large language models to sensitive financial data systems where a CFO's question needs a trusted answer with zero margin for error. Seeing their production-tested solutions against risks like prompt injection highlighted the precise tactics needed to make financial agents safe to ship:
▪️ By design; prevent impersonation with for example user permissions, request scoping on user and session level, MCP token minting, meta data validation & row-level security.
▪️ Proper coding tools for the Agent to post-process balances.
▪️Treat MCP as a regular API; with untrusted callers, token expiry & refresh and user-based access control.
