Anton Kozyrev

Flutter features and releases

New screens and flows from Figma to the App Store and Google Play: state, API integration, push notifications, crash reporting.

Taking over Flutter code

Picking up someone else's codebase, refactoring legacy code without stopping feature work, covering key logic with tests.

Go backend and MVP API

REST API in Go with PostgreSQL and CI on GitHub Actions — enough to launch a product and validate the idea.

LLM features with guardrails

Budget limits, fact-checking and fallbacks: the model does not invent data and does not burn the budget.

Flutter · 4.5+ years

Commercial mobile apps

Platforms for working with external contractors: onboarding, documents, deals, payouts and tax receipts. The code is under NDA, so this is what I did rather than the code itself.

• 10 mobile apps in the stores — from greenfield launches to taking over code from other teams
• Migrated a native Android (Java) app to Flutter at full feature parity and shipped it on iOS and Android; now maintaining version 3.x
• Feature-first Clean Architecture, Riverpod for state, API contracts co-designed with the backend team, real-time chat on Socket.IO, FCM push, Crashlytics, Remote Config
• Key logic covered with unit and integration tests; legacy refactoring without stopping feature work

Flutter · Dart · Riverpod · GoRouter · Dio · Socket.IO · Firebase · Drift/SQLite

Go · LLM · static sites

sitefarm — data-driven site pipeline

A Go pipeline that turns open data into a content site: collect data (NOAA, PRISM) → generate text with an LLM → fact-check with a second model → build a static site.

• Pilot of ~2,100 pages; generation costs about 0.5 RUB (under one US cent) per page
• Numbers come from the data, never from the model; a second model fact-checks every text
• 42 of 50 pages (84%) passed the automated fact check in the pilot run; a rejected page gets one automatic retry, and if it fails again it goes to a manual review queue and is never published
• Budget guard: the model is not called if the worst case would exceed the monthly limit
• Models compared on the real task — a reasoning model spent the whole token limit on reasoning and returned no answer, so it was rejected

Go · Yandex AI Studio (gpt-oss-120b) · Hugo · Cloudflare

Go · PostgreSQL · PWA · LLM

Terroir — personal wine assistant

An assistant with memory: cellar, taste profile, wine matched to your own recipes. Retrieval-grounded answers: the code picks the data, the model only explains it.

• Answers are validated: the model cannot reference a bottle that is not in the cellar
• Monthly token budget and a fallback when the LLM fails
• 90+ tests, CI on GitHub Actions

Go · PostgreSQL/SQLite · PWA · OpenAI-compatible LLM client

B2B SaaS · in development

TapReply — replies to customer reviews

AI replies to reviews from several platforms for venues, in local languages. A Go core with platform adapters: a new review platform is a new adapter, not a rewrite.

• Go core + platform adapters
• Flutter / PWA client
• PostgreSQL

Go · Flutter · PWA · PostgreSQL