Román Tatárnikov

Product Manager | Fintech & Brokerage

Product Manager (Fintech & Trading)

  • Product Ownerpreviously atZeal Group2023 – 2026
  • Product & Risk strategy analystpreviously atOctaFx2020 – 2023
LinkedIn

Name, title and affiliations are given by the model’s author; avva has not verified them. Report this model

Solid coverage

Before your AI advises you on product manager (Fintech & Trading), it reads the decisions Román actually made, the reasons where they recorded one, and where Román would push back.

Use this model for

  • Order execution and routing
  • Margin and leverage controls
  • Pricing and market making
  • Anti-fraud automation
  • Fees, swaps and client credit
  • Requirements and delivery scope

Built from 169 approved decisions across AI chat history. Updated October 25, 2026.

Connect to your AIWorks now, no account needed

Add this link to Claude, ChatGPT, or Cursor. Your AI then checks Román’s decisions before it answers you. The answer is still your AI’s.

https://avva.chat/mcp/product-manager-fintech-trading-roman-tatarnikov
Add Román’s model to your AI

On a phone? Email the link to yourself and add it on a computer.

What your AI sends is used for the answer and not stored. Román is not told.

Try it free
No account, up to 40 requests a day, per model. It returns a smaller part of the model: every stop rule and question, their approach and defaults, the criteria they put first, and five of their decisions.
Full access
Also free: sign in once, then up to 1,000 requests a day, per model. It returns the complete published model, on a connection you can revoke from any device.

See Román’s model answer a question

One question Román answered, and the published decision it rests on. Your situation gets an answer of its own, written by your AI.

Example inside your AI app: a question Román chose, and the answer Román approved

You:

We are building an AI model to catch trading abuse. Should it block accounts automatically from day one, or just flag them?

Checked Román’s Decision Model

Your AI:

Román Tatárnikov would ship it as a detection system first: the model recommends, an anti-fraud analyst makes the call, and no real-time action runs in the first iterations. Before anything is automated, the team measures the model's false-positive and false-negative rates; automatic action comes only once confidence is high enough.

Why. The fear of hurting even one honest client.

What would change this. Measured quality: once the model's confidence is high enough and its error rates are checked, automatic actions can follow.

Outside this model. It does not set the confidence level or the error rates at which automatic action may start, or what the analyst's review must cover.

Copy an example prompt4 patterns with a gap to fill, once you are connected

  • Before you commit

    A full pass over a plan or a draft, with the changes it needs.

    Use Román Tatárnikov's Decision Model on avva before we lock this in. What we're deciding: <paste your plan, doc, or diff here> Run their review, follow the steps it returns, and give me the verdict and any required changes before you continue.

  • Choosing between options

    A choice between two options, their stop rules checked first.

    Use Román Tatárnikov's Decision Model on avva to compare these. Option A: <describe it> Option B: <describe it> Context: <what constrains the choice — margin and leverage controls> Check their stop rules first, then compare the options under their judgment, and tell me which one their model favors and why.

  • A quick gate

    A quick look for anything they would refuse outright.

    Run this past the stop rules in Román Tatárnikov's Decision Model on avva before I send it. <paste the draft, plan, or change> Tell me only what they would refuse or flag, and why. If nothing trips, say so.

  • Before you have an approach

    The questions they ask before an approach gets locked in.

    I'm starting on order execution and routing. Ask Román Tatárnikov's Decision Model on avva what they would want settled first. Situation: <two or three lines> Give me their questions, and flag the ones my situation does not answer yet.

Patterns with a gap to fill. None is an answer Román gave. The one worked answer on this page is the one they approved themselves; anything else here would have been written by us, under their name.

Connect Román’s Decision Model to your AI

Three steps. The first works now, with no key or account; the last makes your AI check Román’s model without being asked.

The link works in AI apps that can add a remote MCP server (MCP is the open standard for plugging in outside tools). You set it up once.

  1. A connector is an address your AI reads from: one line, nothing to install.

    https://avva.chat/mcp/product-manager-fintech-trading-roman-tatarnikov
    1. Open Claude on the web or in the desktop app and go to Customize → Connectors.
    2. Press + next to Connectors and choose Add custom connector.
    3. Paste the address above and name it avva-product-manager-fintech-trading-roman-tatarnikov.
    4. When Claude asks to allow avva, say yes.

    Missing on a work account? Ask your admin to allow it. Claude Code and Cursor take a local config instead.

    The link is added on the web or in the desktop app. Email this line to myself

    The trial is enough to see how this expert reasons and whether their thinking fits your problem. Step 2 returns the rest.

    Check it worked. Send this in the chat once the link is added.

    Ask Román Tatárnikov's Decision Model on avva what kinds of decision it covers, and summarize it for me in three lines.

    Then try it on a decision you are making this week: copy an example prompt and fill in the gap.

Not legal, medical, financial, or other professional advice. Terms of Use

About this model

Every rule below rests on decisions Román approved. Open any of these to check the grounding yourself.

How this model is groundedWhat it holds, and what Román weighs first

This model holds 169 decisions, 26 forces, 13 criteria, and 22 stop rules.

What Román weighs first

Proven on history before it alerts, blocks or shipsMUST WEIGH
One owner and one source of truth for each concernMUST WEIGH
The first release is the highest-impact slice; the rest is stagedMUST WEIGH

The expert set this order. 3 of 13 criteria are shown, the ones the model puts first. The other 10 reach your AI too, along with what each one looks for, the stop rules, and what changes Román’s answer.

Román’s usual answer changes under 12 named conditions. Those conditions and the 22 stop rules reach your AI with the rest.

See examples of approved decisions5 of 167, written from Román’s approved decisions

The first 5 in the record, in published order. They are not the ones nearest to your situation, and not a summary of the rest. Your AI does not copy a decision it reads here: it checks what decided that case against your own facts, and weighs it beside which pressures tended to settle Román’s cases. With a key or signed in, your AI receives all 167, ordered against your situation.

  1. What decided itclients stopped out or margin-called

    Situation: At an MT4/MT5 broker, equity-triggered leverage control cuts a client's leverage automatically and can stop the client out right after the change, while elevated-margin periods also move margin requirements; the call was whether to check the margin outcome before the automatic change.

    Decision: Before an automatic leverage change, calculate the projected margin level and skip the change if the result would fall below a set threshold. A 100% default was proposed at first; later requirements made the threshold configurable.

    Why: It mitigates stop-out risk for clients.

  2. What decided itfalse alarms that hit honest clients

    Situation: A broker was planning AI/ML support for its in-house anti-fraud team — analyzing trade history and client profiles, scoring risk, supporting investigations — and had to settle what the model itself may do in its first iteration: recommend, or act on clients.

    Decision: In the first iteration the model detects and recommends with no real-time action, and an anti-fraud analyst makes the decision; automatic actions come only with sufficient confidence and after quality has been checked.

    Why: The worry is hurting even one honest client.

  3. What decided itone owner and one source of truth · trade-server load and latency

    Situation: Partial fills could be built inside smart order routing or by extending a standalone MT5 gateway to aggregate fills and calculate VWAP.

    Decision: Build partial fills inside smart order routing rather than developing the standalone MT5 gateway, knowingly taking on a broader scope for the current initiative.

    Why: One mechanism serves current and future trading platforms and institutional clients, opens the way to retiring the legacy bridge, and cuts execution latency for clients.

  4. What decided itthe broker's market and abuse exposure

    Situation: Placing a market maker's limit orders on binary contracts, the team had to trade off the reward for posted liquidity, queue priority and the risk of unwanted fills.

    Decision: Do not chase the top of the book at any cost: when a deeper level still earns the liquidity reward with a lower chance of an adverse fill, quote there.

    Why: Being filled does not always earn a reward, while resting deeper in the book still earns the liquidity reward at lower risk.

  5. What decided itwhether the work can still change the decision

    Situation: A broker was designing a web trading terminal to be embedded in a client portal being rebuilt at the same time, and the team proposed studying the audience's reaction to it; the decision was whether to run that study once the product owner's choice was already fixed.

    Decision: Do not run a large study of reactions to the web terminal when its results can no longer change the decision the product owner has made. This is not a ban on user research in general.

    Why: The owner makes the choice in the end, so a study adds no value.

How Román takes up a decision8 steps in order, approved by Román

  1. Step 1Start from what already exists: what is built, what is not and why, which components and data are there, and what depends on them, before planning new work.

    Backed by 6 decisions

  2. Step 2Ask whether the work can still change a decision; a study, an analysis or a paid service that cannot is skipped or switched off.

    Backed by 4 decisions

  3. Step 3Cut the first release to the highest-impact slice and stage the rest, keeping the full requirement written and prioritized.

    Unless: a narrower scope would leave two mechanisms doing one job or an open route around a new rule; then the wider scope is taken knowingly

    Backed by 13 decisions

  4. Step 4Before anything acts on a client's account automatically, project its effect on margin and stop-out risk.

    Unless: the client makes the change themselves

    Backed by 7 decisions

  5. Step 5Validate on history before alerts, automation or a new algorithm go live: false-positive and false-negative rates, other market sessions and volatile periods, the exact providers and time window; the logs, not a smoothed chart, settle what happened.

    Backed by 12 decisions

  6. Step 6Weigh build cost after client risk and ownership: get the engineers' estimate before choosing between options, and keep a costly part in scope with its own estimate rather than cut it.

    Backed by 5 decisions

  7. Step 7Give each concern one owner and one source of truth; when a fix belongs to another team, hand it to them and get their ETA instead of promising one.

    Backed by 12 decisions

  8. Step 8Get the business priority and management approval before shifting other work or locking a scope.

    Backed by 6 decisions

Technical methodologyHow the 104 decisions that say what won and what lost are read
  1. 169decisions Román approved, from AI chat history
  2. 26forces those decisions turned on, the pressures Román decides under
  3. 159contests read off 104 decisions: every force that settled a case is set against every force that lost there, so a case naming two of each counts four times
  4. In groupswhich forces tended to settle cases over which, read off those contests

Nobody ranks these forces by hand. Where any order between them comes from is on For AI agents.

How Román decides, in summaryWho they are, and a summary written from the decisions they approved

Product Manager with 8+ years of experience in fintech, combining strategic thinking, business understanding, customer needs and technical expertise. My career evolved from trading risk management to identifying business opportunities, shaping product strategy and driving product development.

Consult this model on brokerage and trading-platform product decisions — order execution and routing, margin and leverage controls, quoting and market making, anti-fraud automation, fees, swaps and client credit — and on how to scope, stage and validate such initiatives before they ship.

The summary covers what the model holds. Román knows more than is in it.

What’s trustworthy hereWhat is checked, and what is only stated
  • Identity: Name, title and affiliations are given by the model’s author; avva has not verified them. Judge by the record on this page and by what your AI receives.
  • Scope: Product Manager (Fintech & Trading) (Order execution and routing, Margin and leverage controls, Pricing and market making, Anti-fraud automation, Fees, swaps and client credit, Requirements and delivery scope, Trading data infrastructure)
  • Coverage: A solid set of decisions behind it that say why, with patterns that repeat across them.
  • Privacy: Their cases and rules, written in English from decisions they approved. Their own AI takes out client and employer names, and they check every case; source quotes never publish.
  • Whose answer it is: avva runs no AI of its own. Your AI reads this model and answers you, and it names where the model had nothing to say.
  • Limits: Informational tooling for your AI, not professional advice. avva is provided as-is.

Privacy Center · Terms

Know someone whose decisions you would want your AI to have? Send them to Studio. Publishing is free during beta.