Built solo.
Led at scale.

Dating is broken by looks-first swiping. Anketta is a bet on the opposite: text-first, where people are matched by what they actually write and believe — and I own every call, product, business and engineering.

Eight backend services and two apps run on high-availability Kubernetes with GitOps deploys, canary releases and encrypted secrets — 99.5% availability and 15–20 deploys a day, operated by one person.

A self-hosted multilingual model matches people on their manuscripts, not their photos. Signals you leave — resonant, or off-key — reshape the feed live, and every embedding stays local. Privacy isn't a setting here; it's the architecture.

A cascaded moderation pipeline runs cheap checks first and language models only when needed, with cost-tier gating and circuit breakers — so it stays fast and inexpensive at scale.

And it's shipped by a fully autonomous multi-agent pipeline — code to tests to docs to an atomic commit to a merge request — the engineering practice I'd scale to a team for compounding velocity.

Two distributed teams — about sixteen engineers under two leads — were about to fork authentication. One used native REST decorators on a shared Nest.js auth library; the other's non-native GraphQL blocked clean integration, and the two were drifting apart in maintenance and security posture.

I reframed the fight from tech-stack preference to the real cost — maintaining and reusing a shared library. The compromise: the GraphQL team kept GraphQL, but moved onto the native Nest.js library. Both teams landed on a single auth surface in about two weeks.

Then I made both teams' systems more resilient: OpenTelemetry, Grafana and Prometheus across services so on-call could trace incidents across team boundaries; idempotency, retry/backoff and dead-letter queues on Kafka; and RFC review across both leads to stop the stacks diverging again — while still shipping features alongside the ICs.

I spotted a long-standing bug in the credit-metering pipeline: the tracker logged only left-clicks, missing middle-clicks and context-menu opens — common on recommendation widgets. I instrumented production traces to prove it, and roughly 30% of billable engagement was going uncounted. The fix converted that into prepaid credit consumption — real revenue, with no price change to customers.

I led a zero-downtime migration of the high-throughput event-tracking service — the click/impression tracker feeding metering — from a legacy Node service to a refactored TypeScript one. A gradual percentage-based cutover, with full load-balancer rollback on any error-budget breach, shipped it with no customer-visible incidents.

Then, org-wide: I drove eight teams' fragmented services — mixed Meteor, Hapi and Express on mixed Node versions — into one Nest.js monorepo. I authored the migration ADR and aligned eight leads over async memos, cutting the on-call burden of inheriting unfamiliar stacks on rotation.