By Teresa Torres · producttalk.org · @ttorres on X · LinkedIn
This podcast episode features three leaders from Zencity, a company that helps local governments understand residents by combining survey data, 311 calls, social media, and local news into one platform. The conversation explains how the team structured its data in layers, from raw inputs up to briefs, and why that architecture determines what AI can reliably do. It also covers how they pair deterministic systems with LLM-driven synthesis, using citations, evals, and guardrails to protect accuracy and trust. The discussion matters for product builders because it shows how grounding AI in clean, multi-source data and designing around real workflows can make AI outputs actionable in a high-stakes setting like public administration.
01Key takeaways
- Your data architecture largely determines what your AI can credibly do, so invest in it early.
- Layer raw inputs into progressively refined units such as insights and briefs before asking AI to synthesize.
- Pair flexible agentic behavior with deterministic checks to keep outputs trustworthy in high-stakes contexts.
- Use citations and evaluation methods so users can verify AI-generated claims instead of trusting them blindly.
- Design AI outputs around real, recurring workflows so insights arrive when decision-makers need them.
02Key sections
- What Zencity does
- The hosts introduce Zencity as a platform that helps cities reach, understand, and act on community voices. The focus is on turning many messy sources into usable civic insight.
- Layered data model
- The team organizes information in layers, moving from raw data to elements, highlights, insights, and briefs. This structure is presented as the foundation that shapes everything AI can do on top of it.
- Agentic AI with deterministic trust
- The guests discuss balancing flexible, agentic assistants with predictable, rule-based systems. Access to multi-tenant data is negotiated safely through MCP servers and guardrails.
- Evaluating accuracy and latency
- The episode explores how to evaluate AI output when speed matters, using citations, evals, and model-as-judge approaches. Transparency is treated as more important than impressive-looking outputs.
- Workflows and jobs to be done
- Government workflows such as annual budgeting and crisis communication are used to show how AI briefs reach the right people at the right time. The team frames these workflows as a practical jobs-to-be-done lens.
03From the post
“Listen to this episode on: Spotify | Apple Podcasts How do you use AI to help city leaders truly hear their residents? In this episode, Teresa Torres talks with Noa Reikhav (Head of Product), Andrew Therriault (VP of Data Science), and Shota Papiashvili (SVP of R&D) from Zencity, a company”
04Frameworks mentioned
Summary and takeaways written by PM Atlas; quotes are short excerpts. © the original author.