After nine years at T-Bank I left to answer a question I could not test inside a company: how much of a real engineering organization a single person can run when AI agents carry the execution. Not a lab experiment — live products with real users.

The process runs like an engineering organization's: 15+ functional roles played by agents — backend, frontend, SRE, security, design, and QA among them — with review gates between them, ~90 recorded architecture and product decisions, invariants-first specs, and a 14-point definition of done. Using that process I took Routinka, a PWA for routines and habits, from idea to live production single-handedly: architecture, CI with rollback, decision records, runbooks. Calmel, a family logistics service, came next through the same process — what I am testing there is no longer the process but whether anyone will pay.

Everything I now teach about agentic development — to mentees and to teams — was tried here first, on my own project and at my own expense.

T-Bank engineering runs on internal platforms — 440 products in the company registry — and by 2025 the same capabilities were being built in parallel by different teams. Standardizing them became one of four strategic IT projects, and I came in to lead its engineering: a team of six, and a goal sheet where every key result had stood at zero since January.

I started with an honest baseline (~5% real progress), then rebuilt goal-setting end to end: every OKR linked through key result to project to task with nothing left floating, monthly check-ins, clearly assigned product roles. In parallel I built a registry of duplicated capabilities with an FTE estimate against each — the first time the cost of duplication, around 15 FTE, was visible in one place — and moved responsibility for standards off the architecture committee, where decisions had been piling up, into the project itself.

Six months later OKR completion stood at 75% against a 70% target. A service that had been stuck between two teams for two years shipped in nine months, including a full rewrite onto another stack; the company's first internal RAG prototype ran on top of the registry; the registry's marketplace reached 4,000 monthly active users.

I took over the group half-built: three of six planned teams existed, the rest of the engineers were still organized by function rather than product, and three company-scale programs were already on the calendar. I wrote the strategic case — which domains, which teams, what headcount — and got it approved by the CTO.

The group grew from ~30 to 60 engineers across six product teams and seven domains, with hiring and career development built from scratch. The three programs landed in parallel and on time: migrating a partner bank's clients into our app against a hard regulatory deadline (~5,000 clients, under 1% hit corner cases, zero critical incidents), replacing six infrastructure systems — from Oracle to Kafka — under critical-infrastructure law, and making every key service ready for 1M MAU and 100k DAU.

Reliability held at 99.9%+ with 3,000 RPS peaks for two years. In three years none of the 60 left voluntarily and nobody burned out — under constant deadline pressure. Engineers I worked with grew into team leads, one reaching Group Lead and one becoming an architect, with ~30 promotions among my direct reports and ~60 counting the indirect ones.

The hardest wins were the ones with no authority attached: when a partner bank's own app had to be decommissioned, I spent months on their side of the table — motivations, fears, a transparent plan for their team — and closed out the decommissioning with the team intact.

The B2B platform had grown faster than its front door: the teams consuming it had no single place to see what existed, and every question turned into a chat thread that took days to resolve. I pitched a unified developer portal, registered the domain and stood up the first version myself, then staffed it properly — half a team's time, plus an engineer found through internal rotation.

I onboarded the first fifteen-plus platform products personally and set up a support channel with an on-call rotation; later we wired AI into it to answer the recurring questions. The portal reached 2,400 monthly active users and ~40,000 total views; the channel grew to 750 members — about 70% of the business line's IT staff — and resolution time fell from several days to under 24 hours.

The approach outgrew the business line: the portal became the model for the company-wide platform registry I later led as Head of Engineering.

I became the lead of a five-person iOS team within months of joining, at a point where the app was growing fast and releases were still assembled by hand. Over six years the team grew from 5 to 15 engineers — at peak ~30 across three sub-teams — through 100+ interviews and structured hiring.

The engineering practice was built from scratch: Scrum, CI/CD on Fastlane and TeamCity in place of manual releases, trunk-based development, feature flags, a release train that went from ad-hoc to weekly, unit and UI tests where there had been none. I drove the migration from a monolith to a modular monolith — throughput per engineer roughly doubled — and led a cross-team acceleration effort that cut the change lead time from 12 hours to 6–7.

The app shipped every two weeks to 100k+ daily users, with the crash-free rate improved from 99.8% to 99.95%. Along the way ~200,000 lines of legacy Objective-C were rewritten in Swift, and I designed and open-sourced Highway, a unidirectional data flow framework for iOS.