Role. Software Engineer, Modo Labs (10+ years tenure). Owned the admin and configuration experience for the product, working across the PHP application layer, the retrieval and indexing pipeline, and the Kubernetes platform it runs on.

Problem

An LLM knowledge assistant is only as good as what an administrator can feed it and control. The model answers from retrieved documents, so someone has to decide which sources get indexed, how the assistant behaves when a question lands outside its knowledge, and what it is allowed to say. That control surface was the gap. Without a real admin experience, configuring the assistant meant hand-editing settings and re-indexing blind, with no clear way to see what was in the corpus or to set guardrails. The job was to build the configuration layer that turns a retrieval pipeline into a product a non-engineer can run safely.

Approach

  • Built the admin UI and its backing services for managing the document corpus: connecting sources, triggering indexing, and showing administrators what had actually been ingested and when, so configuration stopped being a blind operation.
  • Wired the configuration layer through the full request path, from the PHP admin application down to the retrieval and indexing services, so a setting an admin changed mapped to real behavior in how the assistant retrieved and answered.
  • Handled streaming responses in the admin and preview surfaces, so answers render token by token and administrators can watch the assistant behave with their settings before shipping them.
  • Exposed guardrail configuration in the admin experience: controlling how the assistant responds when retrieval comes back empty or low-confidence, and keeping answers grounded in the indexed corpus rather than the model's open-ended output.
  • Ran large cross-layer refactors to keep the configuration code coherent as the product grew, reducing duplicated settings logic and the number of places a single behavior was defined.
  • Did release-engineering work to ship these changes reliably on a Kubernetes platform shared across multiple product lines, treating deploys and rollbacks as part of the feature, not an afterthought.

Impact

  • Administrators can configure the assistant end to end, from source selection and indexing to guardrails, without touching code or filing an engineering ticket for routine changes.
  • Configuration changes now flow predictably from the admin layer through retrieval and indexing, so what an admin sets is what the assistant does.
  • The refactors left the configuration code with fewer duplicated definitions and clearer ownership, which made later AI features cheaper to add.
  • Releases ship on the shared Kubernetes platform with rollback paths in place, so admin and assistant changes go out without disrupting the other product lines that share it.

Stack

PHP · Kubernetes · LLM / retrieval-augmented generation (RAG) · document indexing and retrieval · streaming responses (server-sent token streaming) · JavaScript · SQL · release engineering / CI/CD

Details are generalized to respect my employer's confidentiality; the specifics are ones I can stand behind in an interview.