Technical Product Manager · ML / AI · R&D

Aleksey Kustov

I'm drawn to places where nothing exists yet — no product, no team, no clarity. The engineering half of me works out what is possible here. The product half works out what is worth doing.

  • +9%

    GPU utilisation — with no hardware bought

  • 10 → 50%

    of R&D teams on the experiment platform in a quarter

  • days → a minute

    data delivery to the consumer at T-Bank

I build products that engineers use: platforms for ML experiments, tooling for research teams, work with data. I wrote code myself for ten years, so I speak to engineers in their language and to the business in deadlines and money. I have been doing product for five years: I started in a startup, where uncertainty is simply part of the job rather than a reason to stop, and for the last three years I have been building internal products at T-Bank and Sber, where scale is added to the same uncertainty.

Aleksey Kustov

Path

One path rather than a change of profession: from code to a team, from a team to a product. Each step added responsibility; the engineering foundation stayed — still a working tool, not a line in a biography.

  1. 2025 — present

    Sber · SberDevices

    Technical Lead · internal ML/R&D platforms

    Sber is the largest bank in Russia; SberDevices is its consumer-AI and devices arm. I build and grow the internal products its ML and R&D teams run on. The main one is the experiment pipeline: researchers launch, track and validate ML experiments in it, and the business sees every experiment and every unit of GPU spend in one place for the first time. I took the product over after a year and a half of stagnation, when it was heading for shutdown: in a month I cleared the product and technical debt, brought the stakeholders back and kept the headcount allocated to it. Then came adoption — demos, working through real pain points, staged onboarding, a public roadmap driven by what teams asked for. I run the product inside a team of thirteen and with no one reporting to me: where strong ML teams do not treat an instruction from above as an argument, only demonstrated usefulness works. The second job is hardware efficiency: I designed and ran a pilot that shifts capacity between production inference and training. Separately I built a multi-agent system for managing releases on GigaChat, Sber's own LLM — the product concept, the architecture together with an architect, and a working prototype.

    10 → 50%

    of R&D teams on the pipeline within a quarter — with no administrative leverage, through demos, real pain points and a public roadmap

    +9%

    GPU utilisation: at quiet hours servers move to training — as if every eleventh one appeared without being bought. One such server, eight H100s, costs $300–500K on the market

  2. 2023 — 2024

    T-Bank

    Technical Product Manager · trigger platform

    T-Bank is one of the largest digital banks in Russia, built without branches. The trigger platform is an internal B2B product giving the bank's business lines event data as a service rather than as a one-off export. I took it over as a proof of concept when there was no team yet: I hired six people and led them by the work rather than by reporting lines. The main thing I did was take the human out of the delivery chain: an analyst used to collect requirements, a partner configured things on their side and sent files, and the team unpicked them by hand. I wired the platform into the bank's other internal products and shipped ML models for computing triggers, getting the move to real-time inference agreed.

    1 quarter

    from proof of concept to a working MVP

    days → a minute

    trigger delivery to the consumer. Everyone who used to receive them by hand moved onto the platform — 100% of it without manual work

  3. 2021 — 2023

    Unikoom

    Product Manager

    An AI travel startup — three teams: data science, backend and mobile. I arrived as mobile engineering lead and grew into the app's product manager. I hired the mobile team from the market and shipped the MVP to production within a quarter — from that point the company tested hypotheses in weeks rather than quarters. Together with data science I took the product's ML core to production — our own route-matching model, trained on a dataset we collected ourselves. Ran discovery and prioritisation: CustDev, A/B tests, RICE. Combined the product role with an engineering one: Flutter architecture, the React web version, CI and environments. We did launch and test a subscription, but never reached a readable picture: users came once and no stable core formed. The product never found product-market fit.

    0 → prod

    hired five engineers and shipped the MVP within a quarter

    ML core

    our own route-matching model — from dataset to production, with data science

  4. 2020

    Vedidev

    Team Lead, web and mobile

    A deliberate move from engineering into product. I owned the MVP of a loyalty app for a retail chain end to end — from business requirements, journey mapping and roadmap down to the React Native code. Coordinated the front-end team, ran technical interviews and onboarding.

    MVP

    owned end to end — from business requirements and CJM to code

    the turn

    from engineering into product: the first role where I answered for more than my own code

  5. 2010 — 2020

    Freelance

    Fullstack developer

    Ten years of client work across JavaScript, C#/.NET and PHP, always with the client directly — which is where the habit of starting from the business problem rather than the technology comes from. The main project was an online learning platform: business logic with the product owner, architecture, front end and back end, shipping and support.

    10 years

    with my hands in the code

    0 → 1

    a learning platform end to end, logic to production

Where I am going

I am aiming at companies where the product runs into infrastructure, data and research teams — where you have to understand the domain, not just run a backlog. My best environment is high uncertainty: the product is still to be invented, and decisions have to be made before all the data arrives. In startups that is called Tuesday.

I came into a large company from ten years of freelancing, where the only thing I answered for was the result — and for a long time I assumed the result would speak for itself. It does not. Inside a large organisation it has to be carried to people, and that is separate work, which I now plan for in advance.

Skills

Product

  • Technical Product Management
  • AI/ML Product Management
  • Product discovery
  • Product strategy
  • Roadmap
  • Customer development
  • A/B testing
  • Customer journey mapping
  • RICE
  • Unit economics

ML and data

  • MLOps
  • ML experiment infrastructure
  • GPU utilisation
  • Real-time inference
  • LLM
  • Multi-agent systems
  • Event data
  • Python
  • SQL

Engineering and process

  • JavaScript / TypeScript
  • React
  • React Native
  • Flutter
  • C# / .NET
  • CI/CD
  • Agile
  • Scrum
  • Kanban
  • Jira
  • Confluence
  • Figma