RAG-grounded LLM workflows that answer enterprise questions from real company knowledge — not a demo corpus.
Peoplebox runs enterprise HR — goals, reviews, surveys, HRIS — and both support teams and customers' employees kept asking the same policy questions. An AI assistant looked like the answer, but only if its answers could be trusted with real company data.
A bare LLM is confidently wrong. Ask it about a company's actual policy and it will invent one. For enterprise customers, an ungrounded answer is worse than no answer — it erodes trust and is a liability.
Built a retrieval layer that chunks and embeds the real knowledge base — policies, docs, past tickets — into a vector store.
Grounded every prompt in retrieved context so the model synthesises sources instead of recalling from memory.
Created an evaluation set that scores retrieval quality before any change ships — the model is only as good as the evidence it is given.
Packaged the flow as a reusable Rails service that other product areas can adopt.
Retrieval quality, not prompt quality, is where the feature lives. We measure grounding before we trust the model — evaluation is the feature.
100%grounded
every answer traced to a retrieved source — never generated from memory
The REST APIs and workflows behind Peoplebox's core HR modules — one service layer powering six product areas.
6 product areas — Goals, Reviews, Surveys, HRIS, IDPs, and reporting on one service layer
EXPLAIN-first schema and query work that kept enterprise dashboards and workflows fast as data grew.
p95 response time — flattened on million-row tables by index-backed, EXPLAIN-first query plans
The reliability layer that keeps employee and identity data in sync across customers' HRIS and identity systems.
24/7 in sync — employee and identity data flowing reliably across customer systems