Projects & case studies · Konstantin Pestekhin

A week → 4–5 hours

Agent harness: from an epic to review-ready code

AI agentsMulti-agent pipelineMCPYouTrackPostgreSQL

Problem: Breaking an epic into tasks took an analyst about a week, and implementation took another 1–2 weeks of development. Tasks were uneven, and pitfalls surfaced only in code.
Solution: I designed a multi-agent pipeline. An agent analyses the epic, studies the codebase to find where the change belongs, asks clarifying questions about logic and modules and surfaces pitfalls, checking hypotheses and the data model against a test database. It then slices the epic into tasks, analyses and tests them, and delivers them ready for YouTrack in a format an AI agent can execute end to end. An agent implements, another agent reviews, and one developer does the final review.
Outcome: Task preparation for an epic takes 4–5 hours instead of an analyst’s week. Implementation including review takes 4–5 hours instead of 1–2 weeks, with a single developer reviewing.

Delivery 60% → 100%

Java 21Spring Boot 3PostgreSQLVirtual threadsFCMKubernetesPrometheusGrafana

Problem: Delivery lived in the memory of a single HTTP request inside a monolith: a pod crash or a deploy lost the whole campaign. No retries, no deduplication, no metrics; worker addresses were hard-coded. About 60% of pushes got through.
Solution: I extracted a dedicated service with a PostgreSQL-backed queue and explicit state machines. A dispatcher turns a campaign into jobs with idempotent inserts; workers on every pod claim jobs in parallel via FOR UPDATE SKIP LOCKED and send on virtual threads. Retries with error classification, a fallback channel, stuck-job sweeping, daily table partitioning. Metrics, dashboards and alerts from day one.
Outcome: Delivery went from 60% to 100%. Losing a node no longer loses pushes, send rate is tunable at runtime, and every campaign shows what was delivered and why anything was not.

Zero-downtime, 2 teams

KubernetesGitLab CI/CDNginxSpring BootJavaTarantool ClusterGrafanaLokiPrometheus

Problem: The whole project had to move to new infrastructure: apps on virtual machines, a Kotlin/Ktor monolith, a standalone Tarantool, and little to no observability.
Solution: I led two backend teams. Moved the apps from VMs to Kubernetes and rewrote services from Ktor to Spring Boot/Java with architectural changes. Set up GitLab pipelines, network access, DNS and routing, and brought up every service. Migrated a standalone Tarantool to Tarantool Cluster. Set up logs, metrics, traces, dashboards and alerting. Shifted traffic seamlessly through an extra Nginx layer.
Outcome: The project went live on the new platform with no downtime for users: horizontal scaling, reproducible deploys, full observability and alerting.