Co-owner and CEO of an advertising agency since 2014.
mike-plotnikov@mail.ruTelegram @mplotnikovgithub.com/mikeplotnikov
Co-owner and CEO of an advertising agency since 2014. Since 2025 I own and ship my own AI products: a mobile app with 100+ live users on RuStore and the App Store, a speech-analytics product in production, a data parser in production. I decide on measurements, not demos: 26 LLMs benchmarked before picking one, cost per analysis known to the ruble. I build, not just manage.
I audit business processes and deploy agents and automation into a company's daily work. Current project: AI process audit and automation discovery for UnionPipe (St. Petersburg). I moved my own agency onto an agentic workflow: monthly client reports, contractor reconciliation, parsers, outreach. Method: written spec, two review waves, decision log, negative results recorded honestly.
I build and ship LLM systems with measurable quality. Model choice by data: a bake-off of 26 LLMs and 8 STT engines on one corpus, blind grading, "fabricated quotes" as the key metric. Data stays inside Russian infrastructure. Agent-driven development: Claude Code and Codex CLI work from my specs, cross-model review waves, releases and incidents are mine. 17 years in advertising, so I understand the customer before they have finished stating the problem.
Speech analytics for sales calls: in production since 2026-09-16; 26 LLMs × 31 recordings ≈ 2,000 analyses and 8 STT engines × 31 = 248 runs before the model choice; the chosen model matches human analysis at 96.8 %, 2 fabricated quotes out of 227, 3.74 RUB per analysis. 5,782 backend and 269 frontend tests in CI.
"Live with Flowers! Florist", bookkeeping and CRM for florists: 100+ live users including corporate accounts, organic daily growth. RuStore and App Store. Scan cost ≈ 4.26 RUB. The 2026-10-12 Yandex data-center outage was diagnosed and closed by swapping the STT engine.
Data parser for a media project: v1.2 in production, full run 69.5 min, 326 of 327 objects, 3,440 cells; 1,164 tests, mutation analysis as a quality gate.
Kapika, e-commerce performance: August 2026 — 1,619 sales (×4.2 vs July), 14.93 M RUB, ROAS 713 %; September — 950 sales, ROAS 539 %. Brand lift 26.9 → 32.9 % in Russia, 18 → 27 % in Kazakhstan.
Zoom brand (Atex) created from scratch: brand identity 2025, acceptance act signed 2025-10-14 without remarks.
artel, an agentic development framework: 10 skills, automatic model selection and failover on rate limits; review-wave log of 346 entries, Claude → Codex delegation log of 318 entries. Review waves caught defects in the previous wave's fixes: 31 → 9 → 4.
Product owner, spec author, architect, reviewer. Code is written by agents (Claude Code, Codex CLI) from my specifications.
business-process audit, discovery of automation and optimization opportunities with AI.
Python 3.12, FastAPI, PostgreSQL 16 with row-level security, React 19. Anonymization before any LLM call, audio never leaves Russia, Russian providers only (Cloud.ru, Yandex AI Studio, GigaChat). Nightly ingestion and encrypted backups with a byte-for-byte restore drill. 2,564 commits since 2026-06-30. Second opinion from a judge model of a different family.
Flutter + FastAPI, on-device document scanning, golden tests, 6 CI pipelines. 1,322 commits since 2026-02-13. Monetization in preparation.
deterministic, no LLM, GitHub Actions as a serverless runtime, Google Sheets as the client's live interface.
my own agentic development framework, built from scratch on principles learned from shipping products: spec → implementation in fresh workers → two cross-family review waves → owner gate. Hooks instead of rules in prose. Compatible with external reviewer agents.
an applied LLM knowledge base built from my own measured runs (DeepSeek, local qwen3): 80 entries, 496 source locators.
a media platform on raising children through sport (vk.ru/prosport_stat, instagram.com/prosport_stat, youtube.com/@Prosport_stat, t.me/prosport_stat): producer — commercial side, media kits, sponsor sales. Season 1 (2026): 10 episodes, 5.45 M monthly reach as of 2026-09-07, ×17 growth over the season.
synthetic interviews with a validity gate; 958 tests, a full rehearsal for $5.50. No commercial launch.
A team of 10+ people, including production. I run client relationships personally: strategy, negotiation, approvals; the team executes.
Sponsorship and special projects. 2012 milestones: Mary Kay, Cinema Park.
Claude Code (skills, subagents, hooks, worktrees), Codex CLI as executor with explicit model and effort, a four-criteria delegation gate, two-wave "author ≠ reviewer" review, model routing recorded as decisions. 45 skills, 22 agents, 10 hooks in the working system.
Own harnesses: bake-offs, blind grading, pre-registered thresholds, Wilson intervals, Fisher's test, bootstrap; mutation testing as a test-quality gate. No eval platforms used (promptfoo, Langfuse).
Full-time, or a project-based AI transformation engagement. Remote or hybrid. Happy to start with a project: process audit → automation plan → implementation with measurements.
Email me: mike-plotnikov@mail.ruMessage on Telegram: @mplotnikov