Product leader & engineer at the intersection of B2B FinTech and AI. Director of Product at Cinimex Group (KZ · RU) by day; shipping AI-native systems and infrastructure on GitHub to stay close to the metal.
AI beyond the hype — what actually happens when AI meets products, teams and business.
- 20 years in B2B FinTech and banking platforms — from Central Bank national payment systems to enterprise AI platforms
- Grew the practice I lead from $0.2M → $15M annual revenue and 9 → 150+ specialists
- 8 national award-winning projects across TAdviser IT Prize, CNews Innovation, FINNEXT, Best ESG, "Project of the Year"
- Payment Hub — large-scale RF cross-border infrastructure (TAdviser IT Prize 2023) · VTB API Platform — 2,000+ internal teams, 700+ APIs (CNews Innovation 2023) · OpsNext — 99% STP across 5.4B annual transactions · Strela — first corporate-deal personalization system in RF (FINNEXT 2025)
- IEEE Senior Member (#99277800) · Patent holder (RU 2012616579 — Cinimex Test Tool, deployed in 6 of the top-10 RF banks) · Author of Digital Bank (Amazon, 2023) and of When Intelligence Became Cheap (Amazon, 2026) — a three-book series on people, companies and trust under cheap AI
| Project | Stack | Why it matters | |
|---|---|---|---|
| 🏛️ | praefectus-ai — your AI sysadmin under contract: an LLM-powered digital steward for Linux VPS administration | Ansible · Python · Bash · LLM contracts · Mermaid · GitHub Actions | You bought a VPS, not a second job. PraefectusAI hires an LLM agent to handle the boring 90% of VPS ops 24/7 — disk cleanup, encrypted backups, security baselines, drift alerts — under a written contract that physically rules out the dumb mistakes (no unfiltered docker prune, no git push --force, no app data touched). Onboarded in 15 minutes, MIT-licensed, battle-tested on a real Hetzner CX23. This is the contract pattern that makes LLM agents safe in production. |
| 🤖 | ai-tab-optimizer — local-first Chrome extension that turns 100+ tabs into structured, searchable knowledge | TS · Chrome MV3 · FastAPI · SQLite · Claude Code / Codex CLI | Solves the 100-tab problem without sending a single tab to a cloud. Survives 1000-tab sweeps; 45-type discriminated-union protocol; per-batch failover between Claude Code and Codex CLI |
| 💸 | bnpl-flexipay-hub — modular white-label Buy Now Pay Later platform | React 18 · Vite · TS · shadcn/ui · Tailwind | Lets a stakeholder click through a full BNPL launch — Seller, Admin, Customer — before anyone writes a backend endpoint. Three role-based React apps under one router, mock-first data layer |
| 🤝 | dealsnext-insight-hub — Lovable-built MVP for B2B deal personalization | TypeScript · Lovable · React · shadcn/ui | Solo re-prototype of the corporate DealsNext platform (FINNEXT 2025 winner) — a Director of Product testing the same thesis with modern AI tooling instead of waiting on a sprint team |
| 🇬🇧 | FluentLoop — single-user English-learning bot that lives in Telegram | Python 3.11 · Telethon · APScheduler · SQLAlchemy · Docker · DeepSeek | English tutor that lives in Telegram and learns my mistakes. Spaced repetition + mistake-pattern detection; 14 epics shipped end-to-end in one autonomous overnight build with Claude Code |
| 🧠 | clawden-ai — always-on personal AI ops platform on Hetzner | Python 3.12 · OpenClaw · LightRAG · MCP · Telegram | An AI ops teammate that pushes signals before I ask — reads work email, surfaces deal-flow changes, plans next actions. DeepSeek classifier → Claude synthesis → LightRAG memory over an Obsidian vault |
| 📨 | maxgram — personal MAX → Telegram bridge: read and answer all MAX chats from Telegram, without installing MAX | Python · WebSocket userbot · SQLite · Docker · Ansible · Hetzner | Reverse-engineered an unsupported messenger so I can answer it from Telegram. Privacy-first (only routing metadata stored); idempotent dedup; supervisor runtime; Ansible-codified ops on Hetzner |
| 🛡️ | ghostroute — router-level VLESS+Reality routing for ASUS Merlin | Shell · iptables · sing-box · xray | Turns a $100 home router into a privacy gateway for the whole household — no apps on phones, no manual VPN toggling. QR profile generator for mobile clients (scan once, import done); read-only web Console with desktop workbenches and mobile triage pages; multi-channel A/B/C data plane that survives reboots and ISP changes |
Plus a long-running side project from 2018: backtestportaa — Java tool for back-testing investing portfolios with technical indicators.
Public, large-scale collaborations across borders:
- Cinimex Group operates across CIS and Southeast Asia with regional offices including a dedicated Kazakhstan office — gives the team a working footprint across multiple regulatory perimeters
- Volkswagen Group AG — joint dealer-network digital ecosystem and IDP platform with the Berlin office (Kubernetes / microservices, 40-person cross-border team)
- Alibaba / AliPay — adoption of ISO 20022 for cross-border payments processed through Russia
- Central Bank of the Russian Federation (Bank of Russia) — designed and built nationwide payment infrastructure, including BESP, Russia’s centralized real-time gross settlement (RTGS) system, and RABIS-NP, the standardized payment-processing platform used across Bank of Russia regional branches.
- Direct payment channels built into Payment Hub: China (UnionPay, CIPS), India (NPCI, RuPay), Vietnam, Bangladesh — covering ~60% of RF cross-border trade with the Asian region
- McKinsey & Company · EPAM Systems — joint delivery on enterprise digital-transformation programs
- Began my career at Exigen Services (now Emergn) — at the time a top-4 outsourcing firm globally, US clients, English-medium delivery
- Start from the user, not the framework. Every project begins with "what does this make easier for the person using it?" — only then does the stack become a question.
- Ship narrow, then expand. A prototype that does one thing well beats a backlog of half-built features. If real users don't pull on a feature, it doesn't get built next.
- Defaults are policy. Cosmetic choices — copy, ordering, what's pre-selected — have product consequences. They get the same care as architecture.
- Documentation is part of the deliverable. Every project ships with a real README, architecture notes, and CI from day one. If a reviewer can't understand the system from the repo, the system isn't done.
- AI in the loop, not in the way. Production tooling, not a demo — see the patterns below.
Patterns I reach for:
- Discriminated-union message protocols — every cross-process message is a member of one TypeScript union; adding one without handling it doesn't compile (45-type union in
ai-tab-optimizer) - Per-batch provider failover — Claude Code → Codex CLI → heuristic, decided per batch by a small policy module; transient rate limits don't kill long autonomous runs
- Three-tier model dispatcher — cheap classification → medium reasoning → expensive synthesis, picked automatically per task (
clawden-ai's OmniRoute) - Stop-and-resume autonomous runs — run state in SQLite with per-tab status, pending queue, partial results, accumulated metadata; survives reload mid-1000-tab sweep
- Knowledge-graph memory — LightRAG over an Obsidian vault, fed continuously by event-driven bridges; replaces "chat history" as long-term context
- Multi-channel data planes with explicit ownership boundaries — no silent rewriting between layers (
ghostroute's A/B/C channels) - MCP-first context engineering — custom MCP servers (
lean-ctx) and tool-search budgets to keep agentic dev sessions within cache windows
Tooling
- AI runtime — Claude Code · Codex CLI · OpenClaw · MCP servers · LightRAG · DeepSeek · Yandex AI Studio · Anthropic / OpenAI APIs
- Niche stack — sing-box · xray · Reality protocol · Hetzner · Telethon · ASUS Merlin
- Languages & web — TypeScript · Python · Java · React · Vite · FastAPI · Tailwind · shadcn/ui · Docker · SQLite · PostgreSQL · Linux
The systems above are the practice; the books are where the argument gets written down. When Intelligence Became Cheap (2026) works through one question at three levels — the individual in The Human Edge, the company in Business in the Age of AI Agents, and verification itself in The Economics of Trust: when producing an answer costs almost nothing, value moves to whoever owns the outcome.
Earlier: Digital Bank (2023), a handbook on banking digital transformation.
The method behind the series is published in full at cheap-intelligence.vercel.app — free, no sign-up:
| Framework | What it does | |
|---|---|---|
| 📜 | Agent contract | States an agent's mandate, its limits, and the audit trail it must leave |
| 🎯 | Eval set | Catches an agent before it fails in front of a customer |
| ✋ | Human review matrix | Draws the line between what a person signs off and what an agent may decide alone |
| 🔧 | Harness builder | Assembles the working context an agent actually runs in |
Around 380 pages and a dozen browser tools, updated far more often than print.
- Currently focused on: building AI-native systems where digital workers own entire operational functions — autonomous teammates that observe, decide, escalate, and report, rather than chat assistants triggered on demand. The goal is measurable business outcomes: faster deal cycles, lower OPEX, and human teams freed for higher-leverage work. Applied to B2B FinTech operations, CBDC and tokenized settlements, and cross-border payment infrastructure. Published expert profile and articles on this thesis at RBC Companies
- IEEE Senior Member (#99277800) · Skolkovo Foundation expert since 2023 · FSI grant jury member
- 10+ peer-reviewed papers indexed at Google Scholar and ORCID
- Yandex Practicum — System Architect program (2025) · Karpov.Courses — ML & System Design · GoPractice — Product Management Simulator · ICAgile Certified Professional
- Languages: Russian (native), English (C1)
- Open to interesting conversations on LinkedIn
This profile is maintained at eiler2005/eiler2005.