Ever wonder why a payroll question can be answered by AI without a human? Rippling’s AI team cracked the code — and you can steal their playbook.
The Simple Version
At its core, Rippling AI is a single language model that talks directly to a unified employee graph. Instead of juggling many sub‑agents, it uses a few declarative “skills” that map to everyday HR tasks.
How It Actually Works
Flat Agent Architecture
The flat agent lives in LangGraph and resolves entities across payroll, HR, and benefits systems in one pass.
Declarative Skills
Skills are simple SOPs written in natural language that the model executes. They replace complex sub‑agent pipelines.
LLM‑Generated SQL over Cached Schema
A single SQL template, cached and reused, powers all data pulls.
Eval‑Driven Development
Instead of guessing, Rippling runs evals until Wilson’s interval gives a confidence bound on performance. This tells them exactly how many repetitions are needed before shipping.
Technical Deep Dive
They built a custom entity‑resolution layer that normalizes employee records from dozens of source systems. Caching lets follow‑up queries reuse the same schema without hitting the database repeatedly.
Statistical confidence intervals guide the number of eval runs, turning a vague “feel‑good” test into a rigorously quantified metric.
The cost‑uncertainty‑lag triangle forces trade‑offs: you can prioritize speed, accuracy, or cost, but not all three at once.
Smoke evals run on every commit, while health evals gate production releases, ensuring only well‑tested models ship.
Synthetic test environments simulate real HR queries, catching hallucinations before they reach users.
Why This Matters
HR teams get instant, accurate answers, cutting ticket resolution time by orders of magnitude. Engineers save weeks of tooling work by reusing a single SQL generator. The eval framework makes AI releases as predictable as traditional software.
Key Takeaways
- Flat agents can replace multi‑agent architectures for simpler, faster AI.
- LLM‑generated SQL over a cached schema supersedes many bespoke tools.
- Statistical confidence intervals (Wilson’s interval) quantify eval needs.
- Cache data to enable iterative, low‑latency follow‑up queries.
- Smoke evals catch regressions on every commit.
- Health evals ensure production safety before launch.
- Synthetic tests expose hallucinations early.