
A five-part, evidence-led journey from fragmented warranty operations through architecture, traceable RAG, governed human approval, and an honest production-readiness reflection.
Servexa Warranty AI began with a practical operations problem: warranty decisions rarely live in one screen or one document. Support staff and technicians may need to reconcile repair history, product data, policy text, manuals, evidence, permissions, and the current state of a case before they can decide what should happen next. I treated that fragmentation as a systems problem rather than an invitation to add a generic chatbot.
This series explains the major engineering decisions I made as Servexa's primary engineer. I designed and implemented most of the platform across the React application, Express business backend, FastAPI and LangGraph runtime, PostgreSQL and pgvector retrieval path, Redis coordination layer, and human-in-the-loop workflow. The repository includes limited collaborator contributions, which I do not present as my individual work. Likewise, I do not describe architecture documents or planned controls as if they were already operating in production.
Every article separates four kinds of statements:
| Category | What it means here |
|---|---|
| Implemented capability | Behavior visible in committed source at revision 6aca2a1. |
| Test evidence | Behavior exercised by a named, focused suite. It is not a claim that the full system passed end to end. |
| Expected product value | The operational benefit the design is intended to enable, without invented user or business metrics. |
| Verified production outcome | Used only when deployment evidence supports it. This draft series does not claim production scale or readiness. |
On 14 August 2026, I reran focused tests for the HITL service and action registry, shared HITL contracts, and two web approval behaviors: 13 tests passed across six files. Python, full browser, restart-recovery, load, and complete end-to-end suites were not part of that run.
This is not a technology tour. Each post focuses on a decision: the constraint behind it, the alternative I rejected, the implementation I can demonstrate, and the failure mode or debt I would address next. Together, the posts show how I reason about business authority, AI uncertainty, evidence, state transitions, and production risk.
The series is also deliberately a set of reviewable drafts. Its Notion status remains In progress, no publication date has been assigned, and any claim that cannot be defended against the committed repository should be removed or rewritten before release.
The public source is Servexa Warranty AI at revision 6aca2a1. Later uncommitted production-readiness material is discussed only as work in progress in the final article.
The previously listed demo URL returned HTTP 404 during verification on 14 August 2026, so this draft does not present it as a working demo.