Skip to main content

The Uvarov Review

Vol. III · No. VIII

London · Saturday, 22 August 2026

ym8 — “your mate”, London text-speak  ·  A late edition  ·  Founded MMXXVI in London  ·  Vol. III, No. VIII

  • FCA-approvedEMD Individual
  • ECOMMPAY · KYB90 → 12 d — reduced from 90 days to 12 days
  • Altery · KYC approval3.45% → 10.5% — increased from 3.45 percent to 10.5 percent
  • Otkritie · SME book₽70B → ₽130B — grew from 70 billion to 130 billion roubles
  • Moscow Exchange · saved£1.7M / yr
  • Altery · agents in production30+

Lead · The Beat of London Fintech

Vasiliy Uvarov — AI builder & fintech operator, London. Chief Data & AI Officer at Altery. Daily Claude Code + Codex operator.

Chief Data & AI Officer. Still the One Writing the Agents.

Multi-agent orchestration in production, at a regulated UK institution — 30+ agents and 230+ skills live across 7 departments. The operating constraint is practical: production agents need explicit human ownership, review gates, and an audit trail.

By Vasiliy Uvarov  ·  London  ·  Saturday, 22 August 2026

Production multi-agent orchestration is the day job: 230+ skills across 14 domains, 30+ agents live across 7 departments, 550+ API routes across 40+ services. Separately, Vasiliy Uvarov is an FCA-approved EMD Individual; that standing concerns management of an electronic-money institution generally, not AI-specific accreditation or accountability. The orchestration layer was written here, not bought, and the flagship services on top are tested like services rather than prompted like demos. Behind that, twenty years of fintech delivery, engineer to Chief Data & AI Officer. The same habit ships product nights and weekends: three solo-shipped builds — Cosmic CopilotVerify: Cosmic Copilot, live in production (opens in a new tab), AEO Platform, and Prime Legacy — and a receipt measured 21 May 2026: 907,372 lines, 1,959 commits, 7 repositories. Methodology on the README (opens in a new tab)Verify: the MIT-licensed source repository on GitHub (opens in a new tab).

Marks a claim that can be verified against a public record, without an introduction.

Before Altery there was ECOMMPAY, where I cut B2B onboarding from ninety days to twelve, and Moscow Exchange, where a fifty-percent rip-and-replace unlocked £1.7 million in annual savings. Earlier still, Bank Otkritie, where I scaled SME lending from ₽70 billion to ₽130 billion, deployed a hundred-plus RPA bots, and ran a ten-figure IT portfolio across twenty-one engineering teams and two hundred and thirty staff.

The thread is the same in every job: regulated complexity goes in, a shipped product comes out. The change since June 2025 is that the shipped product now contains agents.

“I ship AI as code, not slideware. 100,000+ ai-assisted lines added within one calendar month — and 463 tests green across the two flagship agent services, not a vibe check.”

The three side projects are the calibration. Cosmic Copilot is the same product built twice by one author — native iOS and Next.js web, 676,461 lines measured 21 May 2026, 1,014 commits across the native app and web app in the trailing 90-day window ended 21 May 2026, its readings computed on a deterministic ephemeris rather than generated. AEO Platform tracks brand visibility across answer engines and detects citation drift. Prime Legacy is a real-estate site stitched to a Python parser pipeline.

One bet runs through all of it: that the same person can write the regulatory memo, the product spec, and the code that ships behind them. Every role above is that bet placed again.

Continued on page two: Records of the House · Commissions · Dispatch (22 Aug 2026)

Column · The Builder’s Bench

The Operators Who Still Write the Code.

On why one fintech CAIDO refuses to put down the keyboard, and what that does to the AI transformation underneath.

An occasional column  ·  By the Editor

The operators who still ship code are the operators who still understand the cost of shipping it. 20+ Custom GPTs and 9 Agent SDK apps now sit between Altery and its customers; somebody on the executive floor has to know what each one costsin tokens, latency, and regret. The cheapest way to know is to have written one. Hence the bench — three products, nights and weekends, as a deliberate calibration habit. If a feature is slow to ship on the bench, it will be slow to ship at the firm.

By the Wire · On the AI Bench

The Vibe-Coding Years.

A frank account of the rig. Claude Code as primary IDE, Codex CLI as second opinion, Cursor for hot edits, MCP servers wired in by hand.

An occasional column  ·  By the Editor

He runs this stack daily, by choice, not because the day job requires it. Default model split: Opus for planning, Sonnet for sub-agents, Haiku for mechanical fan-out. A custom ai-coding-agents-* skill family — sessions, sandbox, permissions, runtime, observability, evals — because prompts alone don’t ship. 20+ Custom GPTs and 9 Agent SDK applications live inside Altery today. 100,000+ ai-assisted lines added within one calendar month, derived from the 1 Jan 2026–21 Aug 2026 activity record. Dispatch is a protocol, not a habit: a dependency-aware task graph before launch, frozen interfaces and a verifier command per task, waves that launch only unblocked work, and worker reports rejected when the declared deliverable or the verifier output is missing. The rig works because the rig was built— and because the flagship Altery services under it are tested like services, not prompted like demos: 170 tests on the support service that replaced a seventy-eight-node no-code workflow, 293on the fraud and PII detection engine — including parallel-versus-single-pass equivalence cases, the ones that catch an agent that is fast and wrong. Both suites green. That is an evaluation harness over the retrieval path, not a vibe check.

For Hiring & Referrals

What role is Vasiliy Uvarov looking for?

Vasiliy Uvarov is open to engineering, product, or technical-programme roles on agent platforms, developer tooling, and applied AI deployment. He is based in London on a UK Global Talent Visa and does not require sponsorship. The one-page brief is written to be forwarded internally.

For Advisory Work

What advisory work does Vasiliy Uvarov take?

Vasiliy Uvarov takes a small number of regulated-fintech commissions alongside the day job. The work is operator-led: diagnosing a stalled product, onboarding, data, or agent-system decision and leaving an executable plan rather than a slide deck. Scope and boundaries are on the commissions page.

Frequently Asked · About the Editor

About Vasiliy Uvarov.

10 questions, asked often, answered here on the record.

Who is Vasiliy Uvarov?
Vasiliy Uvarov is Chief Data & AI Officer at Altery, a UK Electronic Money Institution based in London, and an FCA-approved EMD Individual. He builds production agentic AI systems in regulated fintech and ships solo side projects in his off-hours. He was promoted from CPO in June 2025 and has 20+ years of engineering and product delivery experience.
What is his current role?
Chief Data & AI Officer (CAIDO) at Altery, a UK EMI in London. He leads AI transformation across 7 departments, with 30+ production domain agents live, 230+ production skills across 14 domains behind 8 routers, and 20+ Custom GPTs plus 9 Agent SDK applications in production.
Is the multi-agent work real production engineering, or internal tooling?
Both, and the distinction matters. The agents run inside a live UK Electronic Money Institution serving real customers, not a lab or a sandbox — but they are internal-facing systems, not a product sold to third parties. What makes them production engineering is the test posture and the accountability: the support service that replaced a 78-node no-code workflow carries 170 tests including golden-case evals, and the fraud and PII detection engine carries 293, including parallel-versus-single-pass equivalence tests that catch an agent being fast and wrong. Both suites green. The services ship with OpenTelemetry, Prometheus, Vault-managed secrets, and PII scrubbing, and a CI gate fails the build when a legal agent's anti-hallucination disclosure block is missing. Agent failures get postmortems against the MAST multi-agent failure taxonomy rather than ad-hoc prompt tweaks.
Does experience in regulated fintech transfer to a frontier AI lab?
The honest answer is that some of it does not. EMI-specific regulatory knowledge — safeguarding rules, EMD Individual obligations, FCA reporting — is not what a research lab needs. What transfers is the engineering discipline created by deploying agents where errors have consequences: verification infrastructure, evaluation harnesses, refuting verifiers, structural rejection of unverified output, and audit trails as defaults rather than afterthoughts. Verification at volume — making a fleet of agents produce work that survives sign-off, at a cost that scales — is the same problem labs face when deploying agents to customers, arrived at from the deployment side rather than the research side.
Is he an engineer or an executive?
Currently both, which is unusual and worth being precise about rather than claiming to be a full-time IC. The seat is Chief Data & AI Officer at a UK EMI: the day job is strategy, org design, and accountability across 7 departments. The code is real and continuous alongside it — 907,372 lines and 1,959 commits across 7 solo repositories, measured 21 May 2026, single-author on every one, plus the 230+ skills and 30+ agents behind the day job's orchestration layer. The reason for keeping both is stated plainly on the site: an executive deciding what agent systems cost in tokens, latency, and regret should have written one. Anyone hiring for a pure research role should read this as a deployment-and-systems profile, not a research one.
What is his AI coding stack?
Claude Code as primary IDE (Opus for planning, Sonnet for sub-agents, Haiku for fan-out), OpenAI Codex CLI for second opinion, Cursor and v0 for hot edits, and custom MCP servers for tool integration. The measured monthly floor is 100,000+ ai-assisted lines added within one calendar month. Separately, 1,400+ top-level claude code and codex transcript files across the full measurement window, measured 1 Jan 2026 through 21 Aug 2026. Sub-agent work is dispatched under a documented protocol rather than ad hoc: dependency-aware task graph before launch · frozen interfaces and per-task verifier commands · wave-based launch of unblocked tasks only · structural verification of worker reports. The skills and agents library is mirrored publicly under MIT.
Is he FCA-approved, and can that be checked?
Yes, and the second half of the question deserves a precise answer. He is an FCA-approved EMD Individual at Altery: the FCA assessed the appointment, raised no concerns, and added him to the firm's record. The appointment was assessed against the Chief Product Officer seat in April 2024; he moved to Chief Data & AI Officer at the same firm in June 2025 and the standing carried across. Approvals under the Electronic Money Regulations 2011 work by non-objection and are recorded against the authorised firm, so — unlike a senior-manager approval under SMCR — there is no individual entry in the FS Register to search. The FCA's notification is held by the firm's compliance function, and the firm reference is supplied on request. The status concerns management of the regulated firm; it is not an AI qualification and does not approve external advisory work.
What products has he shipped solo?
Three production products: Cosmic Copilot (Cosmic Copilot native app + web app, 676,461 LOC measured 21 May 2026, 5,000+ registered users; paying subscribers exist), AEO Platform (134,178 LOC measured 21 May 2026), and Prime Legacy Homes (30,002 LOC measured 21 May 2026). The separate solo build receipt totals 907,372 lines and 1,959 commits across 7 production repositories, measured 21 May 2026. Methodology: find + wc -l excluding node_modules, build output, Pods, and lockfiles.
Where is he based?
London, United Kingdom. He holds the UK Global Talent Visa.
How do I contact him?
Email vasilyu@gmail.com, or via LinkedIn (linkedin.com/in/vasiliy-uvarov), GitHub (github.com/vasilyu1983), or X (x.com/vasilyu). He takes a small number of advisory commissions in regulated fintech alongside the day job — see the Commissions section.