Hanover Park Innovation & Technology Culture

What’s your rule for releasing fast — and what KPI proves it works?

Merging is releasing. We deleted the release, no trains, no freeze windows. The main KPI is production and the unit of shipping is a stacked difference small enough to actually read. You ship a branch or three, then stack the next ones on top before the first is reviewed. Nobody sits blocked behind their own work.

Speed stays safe because the ceremony scales with blast radius. Low-risk visual changes ship on an adversarial AI audit alone. Features get AI review plus one human approval. Anything touching money gets a domain owner's deep read, every time, no exceptions. We spend rigor where mistakes are expensive and go flat-out everywhere else. If you need a release train to feel safe, you'll hate it here.

The KPI is land rate, straight from git. Other metrics include sixteen pull requests per engineer per week; median difference of 168 lines; median open-to-merge of two hours; weekly land rate is up two and a half times in two months, while the review bar got stricter. Small differences and matched rigor buy speed and safety at the same time.

Which standard or metric defines “quality” in your stack?

We do fund administration. Our output is other people's books: capital accounts, allocations, the statements Limited Partners make real decisions on. Test coverage is a weak standard for that. Coverage measures effort; we measure truth.

The standard is a verification ladder every engineer here knows cold. Typecheck proves shape, tests prove intent, the running system proves behavior and for money none of that is enough. Money needs an oracle: an independent path that computes the same number, or an invariant that cannot break. If no oracle exists for the number you're touching, building one is part of your task. Then decorrelated review: fresh eyes, human or adversarial AI, hunt for what the author can't see.

None of it lives in a Wiki. Canaries assert structural invariants across the ledger and they check for exactly zero dollars of drift. No tolerances.

Name one recent AI/automation shipped and its impact on the team or business.

Email agents. In fund administration the inbox is the API of the business: bills, wire notices, schedule-of-investment updates, Limited Partner contact changes. All of it lands as unstructured email that accountants used to triage by hand. Last quarter we shipped sandboxed LLM agents that read a firm's email, safely query that firm's live data as needed and draft the accounting task ready for one-click review.

We've built out the tools to allow agents to safely draft actions for our accountants to action on. They reach production data through a read-only Postgres role that cannot bypass row-level security: arbitrary SQL in, zero writes, zero cross-tenant reach. The database enforces that, not a system prompt.

Autonomy is earned. Every draft lands in front of an accountant and we manage one number, the no-edit rate — drafts approved without a single change. Accountants now review structured drafts instead of doing data entry and agents earn autonomy the same way engineers here do — with evidence.