A major Czech bank

As an external consultant via ADF, I built a major Czech bank's GitHub Copilot / AIX enablement kit — reusable skills, custom agents and prompts across ~10 repos, with banking-safety guardrails (auth / payment / key-management off-limits, anti-fabrication) and DORA-aware PR review. A reproducible eval harness — golden tasks mined from their own repos, A/B'd across models — proved the kit makes even weaker models refuse-and-escalate on risky changes instead of waving them through.

The engagement

This was an external AI-consulting engagement, delivered within ADF, for a major Czech bank. I was not on their internal dev team and I did not build their banking product. The goal was narrower and harder: enable the bank's own developers to adopt agentic coding inside a regulated codebase, without the AI becoming a new class of risk.

One constraint shaped every decision. This is regulated code, so authentication, payments and key management are off-limits to an agent, and a model that confidently edits a secret-rotation path is worse than no model at all. The brief was never "make the AI faster" — it was "make the AI trustworthy enough that the bank's reviewers can let its output through."

My job was to analyse, not to ship their product. I deep-scanned their regulated front-end platform — built by other vendors — to ground the kit in their real conventions, and I contributed zero lines to it. What I owned outright was the enablement layer: the kit, the eval harness, and a hands-on AI-dev workshop.

The enablement kit

I built the bank's GitHub Copilot / AIX enablement kit across roughly ten repositories: AGENTS.md conventions, a layer of shared skills plus repo-specific ones, prompts, and a small set of custom agents. The point was reusability — one set of conventions a developer carries from repo to repo, not ten local dialects.

The safety surface is where the real work went. I encoded banking-safety guardrails as do-not-touch catalogs for auth, payments and key management, plus anti-fabrication rules so an agent says "I don't know" instead of inventing an API, a config value, or a compliance claim. PR review was wired to be DORA-aware, so the policy reasons about change risk rather than just diff size.

Concretely the kit ships dozens of reusable skills, a set of custom agents and a library of prompts, fronted by an AI-dev workshop with written guides and live demos — so the team leaves able to run and extend it without me.

The eval harness

None of the above is worth much if you can't measure it, so I built a reproducible eval harness. I mined golden tasks straight from the bank's own repositories, had each candidate output blind-judged by an LLM against a five-criterion rubric with two hard safety gates, and A/B'd the runs across a strong model (Opus) and a weak one (Haiku).

With the kit applied, measured quality moved from 78.5% to 95.2%. But the number is not the headline. On a dormant SSO library, the un-kitted model declared a secret-rotation change "merge-ready"; the kitted agent refused the same change and escalated it. The kit made even a weaker model safer on regulated code — it learned to refuse and escalate instead of waving risky work through.

Because it's reproducible, the harness doubles as a regression test: re-run it on every kit change and the quality and safety claims stay honest, not a one-time slide.