Engineering practice
Demos show what a system does on a good day. This page shows how the decisions behind it were made — and what they admit they do not cover.
Decisions are written down
337 architecture decision records across 25 repositories, each written in English and Chinese, each stating the context, the decision, the consequences, and the tradeoffs accepted. Every product carries them. When you ask why a system behaves the way it does, the answer is dated and written, not reconstructed from memory.
Most live in private product repositories. The three excerpts below are quoted verbatim rather than linked.
One answer, computed in one place
The worker can generate one recommendation, the request path can recompute a slightly different one, and the frontend can infer a third from raw fields. If that happens, the product becomes hard to debug and easy to mistrust.
ADR-012, Decision Cache Boundary — Invest AI, 2026-05-20
The rule that followed: decision calculations happen only in the worker. The API may expose a cached decision and report it as stale, but may not create, alter, or repair one on the request path. A financial recommendation that changes depending on which screen you opened is not a display bug.
Enforcement that states its own limits
An agent firewall is only as good as its interception point. If the agent can reach the shell, filesystem, or network without passing through the policy layer, the firewall is decoration.
ADR-001, Interception Point — Agent Firewall, 2026-07-05
Three architectures were compared against real constraints, one was chosen, and the gap was documented rather than left for a customer to find: v1 protects against a tricked agent, not a malicious local user. That sentence is in the README, not only in the ADR.
Saying what a green check does not mean
It cannot tell you that the metric is the right one, that the dataset is representative, or that the sample is large enough for the difference to be real.
Eval Registry — on its own regression gate
An evaluation gate that blocks a release is only as trustworthy as the metric behind it. Stating that plainly, next to the feature itself, is the difference between a benchmark and a number that gets quoted back at you in a procurement review.
What you can read in full
Product repositories are private, but the engineering surface that matters for evaluating us is open:
- xingai-eval-registry — the evaluation registry and its regression gate, including the limits quoted above
- xingai-enterprise-ai-pocs — enterprise proof-of-concepts — MCP gateways, guardrail monitoring, claims workflow redesign
- xingai-enterprise-ai-design — 18 written articles on agent governance, MCP in production, and loop engineering
If this is the level you need
The useful conversation is usually a specific one: an agent you are about to put in front of customers, an MCP surface nobody has audited, a decision path that gives different answers on different screens. Bring that, not a brief.