Case Studies

Extending the Decision Engine from real-time to batch — the hardest part. The bank's entire customer base comes in as input: every customer has to be checked, refreshed, and pushed through a correct data pipeline before an offer can be built. Before me, a full client-wave calculation took up to a month — scattered manual recalculations, ~150 business filters applied by hand, an unmaintainable monolith. In parallel — rewriting a huge legacy system with colossally complex algorithms across the whole customer journey. And all of it on an extremely hostile market with a high key rate.

Leading the cluster: portfolio, resources, strategy defended at C-level. Rebuilt the calculation into a single automated pipeline: Apache Flink pulls data from any source → S3 → 150+ business filters applied automatically → Kafka → offer calculator → prioritization. Real-time and batch on one engine. Over 2025, together with the teams, I fully rewrote the legacy: not a "technical migration" but new, genuinely complex data-processing algorithms across the customer journey.

The full client-base calculation dropped from ~a month of manual recalculation to ~2 days; 150+ filters now run automatically instead of by hand. Time-to-Market grew severalfold — a managed production process in place of an unmaintainable monolith. Billion-scale revenue.

We needed a solution on par with SAS RTDM to push personalized marketing offers to customers in real time with minimal latency. Zero to production in 5–6 months, engineered for high load and scaling from day one.

Built the expert team from scratch in two months. Designed the architecture, defended it at the internal architecture review board, and worked out the scaling scenarios. Delivered a product of 15+ microservices on Apache Flink, Tarantool EE 2.0, Java 21, Scala, PostgreSQL, Apache Kafka, REST API, Istio, Spring Boot 3.

In production since April 2024. First team in the bank running fully on the import-substituted "Sfera" CI/CD. Modern standards throughout: autoscaling, failure resilience, full automation, self-healing. Zero technical debt across the team. Joint talk "Moving a Banking Product to Real Time" at HighLoad++ Saint, June 2024.

Not a ready-made team but a group of people assembled from the fragments of different teams: graph platform, geo-platform, NER services, a data-labeling service for ML. It had to be forged into a team, given a production process, and launched — against a backdrop of migrated legacy, technical debt, fragmented CI/CD, and scattered standards.

Forged the fragments into a working team and set up the process from scratch. Launched the CI/CD migration onto a single pipeline (TeamCity), cleared the migrations and debt, and built out resource management. Kicked off a community of technical experts across the T1 holding — the start of the Tech Guilds.

Automation up from 40% to 80% within the information system. Cut technical debt almost in half, and cleared the critical information-security and support issues.

Turning the VTB project into a full-fledged product, Mirion: it needed full-stack analytics from ETL prototyping through to ML/AI solutions and graph analysis.

Prototyped ETL on PySpark with Spark Catalyst optimization, trained ML model prototypes on Spark MLlib (Spark NLP, spaCy, ruBERT), applied Random Forest and GAN. R&D on graph models (PageRank, Community Detection in GraphFrames, NetworkX) and the customer "golden record" concept. From May 2022 — head of the analysts group, reviewing ETL from ODS → DDS → CDM → ArangoDB graph.

From February 2023 — acting IT lead, resource management included. In parallel — a 9-month IT team management program from Stratoplan (Hi-PO).

Architecture practices lived in silos. Teams kept reinventing the same solutions across different products, technical debt kept growing, and there were no shared standards.

Initiated and launched the Tech Guilds community: led 20+ experts, defended the project at the CEO-1 level, and built shared standards and cross-team knowledge exchange across the holding.

1000+ cases solved by the community. A phased 40% cut in development cost. Two to three weeks of development saved per team on average.

The holding’s data architecture (reference and master data) was fragmented. Daily reporting to US investors in the parent group was breaking down.

Designed the data architecture and defended an adapted Data Governance project before top management — my first hands-on project management. Rebuilt the data flows on a waterfall model within the parent company. R&D on local ML models (pymorphy2, K-Means).

The solution scaled across the entire holding in Russia. The daily-reporting issues for US investors were resolved and operational risk dropped. Stack: Python, SQL, Pandas, NumPy, Seaborn, MS Visio, 1C (DO, KORP, ZUP, UPP, UT).

Running a portfolio and building several products at once usually takes several teams. I don't have them — just me and an LLM stack. I needed a methodology where one person plus AI reliably covers the workload of several teams.

Codified a multi-agent loop where the human sets the vector rather than writes the code: RAG over architecture decisions, analytics over trace logs, auto-drafted ADRs, AI-augmented backlog decomposition. The human's role is like in gradient descent — point the direction, sharpen the thought, get finished analytics, and make the decision on top of it.

By my estimate, one framework with a single Max-tier subscription covers the workload of ~2–3 development teams — with an active human setting the vector. Several products are built on LAF: Dev Planning, a market app in testing, and LAF itself.

Engineers create value but rarely capture it: fixed salaries, opaque options, a burnout culture. The cooperative model is almost absent from Russian IT.

Technology advisor to a federated-cooperation project: shared ownership of development output, transparent stakes tied to real contribution, democratic voting. A Telegram Mini App for governance is in the works.

The platform is in active development with its first wave of members. Inspired by Mondragon, ESOP, and RAD-KOP.

Real-world cases on real-time, data, and engineering management are poorly documented in the Russian-speaking community — every team reinvents the same mistakes.

10+ talks: HighLoad++ Saint 2024 ("Moving a Banking Product to Real Time"), Analysis & PM Conference 2024 ("Building an Expert Team", "Accelerating Development with AI"), Analyst Days 15 (2023, "Behold, Padawan: the Path of Data"), Impulse 2023 ("Community Synergy"), Analyst Days 14 (2022, "Don’t Paint the Grass — data risks in a BANI world").

A track record at major IT conferences — public cases instead of success-only stories. The HL++ Saint 2024 talk is published on the HighLoad Channel on YouTube.

A problem, not a "challenge": you have to design the team's entire workload, assess the production pipelines, and give a real estimate of timelines — without manual project management. And not just estimate, but manage that estimate flexibly. Existing tooling is built for the scrum master, not the tech lead: utilization, slip risk, dependencies — all of it by hand.

Technology partner on the tool: planning and measuring team utilization with an AI module. Automatic backlog parsing, dependency estimation, deadline-slip forecasting, AI-augmented task decomposition, and flexible re-estimation along the way.

The MVP is in active development, with a prototype on a small group of tech leads. The bar isn’t a classic PM tool — it’s manageable efficiency of the production process.