Joe Khosbayar

Backend engineer on payments infrastructure at Capital One. I work on the systems that must not drop anything.

Centreville, Virginia · sending since 2022

100,000+ transactions / day sustained through a live PostgreSQL repartition
250M+ customer profiles moved by a distributed-locking ETL pipeline
95% P95 latency reduction on high-traffic financial transactions
90% less failover downtime via Aurora Smart JDBC driver switchover

Capital One

Side A · June 2022 — present · four positions, one track

Four years at one company, from a first internship to Senior Associate Software Engineer on payments infrastructure — an intern twice, then hired into Consumer Identity, then promoted into Bank Tech a year later. The systems below move money and authenticate the people moving it, so the interesting constraint is always the same one: it has to keep working while you change it.

Aug 2025 Present
Senior Associate Software Engineer

Backend · Payments, Bank Tech · McLean, VA
- Orchestrated a zero-downtime migration to a partitioned PostgreSQL architecture, with automated archival and pruning to hold performance at 100,000+ transactions per day.
- Led the end-to-end migration of high-throughput distributed infrastructure from Kafka to MSK, with zero latency impact for the Debit vertical.
- Resolved critical PostgreSQL bottlenecks in distributed data systems, cutting P95 latency 95% on high-traffic financial transactions.
- Automated integration testing for payment services, accelerating production delivery 2× while holding correctness.
- Built an automated provisioning and rehydration pipeline for Redis Enterprise clusters.
- Implemented fault-tolerant database switchover on Aurora Smart JDBC drivers, reducing downtime during failovers by 90%.
Aug 2024 Aug 2025
Associate Software Engineer

Backend · Authentication, Consumer Identity · McLean, VA
- Architected a resilient ETL pipeline over 250M+ customer profiles using distributed locking and asynchronous processing.
- Deepened observability across logging, monitoring, and alerting to support system reliability.
- Redesigned the integration testing framework for a WebSocket service, improving reliability and correctness.
- Implemented blue-green deployments for zero-downtime releases and fault-tolerant rollbacks.
Jun 2023 Aug 2023
Software Engineering Intern

Enterprise Consumer Products · San Francisco, CA
- Built a Notification Service for internal teams.
- Developed a REST API letting client teams configure event-based notifications.
- Shipped a Lambda with Slack integrations, and the front-end component for the first client — A/B testing status updates.
Jun 2022 Aug 2022
Software Engineering Intern

Enterprise Consumer Products · McLean, VA
- Built an internal platform for Risk Associates to manage QA control automations.

Mighty

Side B · Sole engineer · Dec 2025 — present

A real-time multiplayer server for Mighty, the Korean trick-taking card game. Live at themighty.gg.

themighty.gg source

Two gigabytes of RAM and a $10 / month budget decided everything.

The whole stack — game server, database, cache, proxy and telemetry collector — runs on one 2 GB ARM box for about ten dollars a month. That budget puts its telemetry on a free tier that silently drops data past 10,000 metric series, failing precisely when monitoring is needed. Everything below is built inside those limits.

Six containers on one 2 GB ARM box. Only the proxy publishes a port, and only the collector talks to the outside.

Running in production

7 systems · live at themighty.gg

Authenticated real-time play: Full game sessions held open over WebSockets behind Cognito-backed sign-in, with an authentication timeout on the handshake, an origin allowlist, and per-user and per-IP connection caps.
Rate limiting at two layers: Ten messages a second with a burst of twenty on every socket, plus a Redis-backed HTTP limiter behind a custom proxy build. Both are instrumented, so whether a real player has ever been clamped is a question with an answer.
Version-checked game state: Every mutation is optimistically concurrency-controlled against a version, and the conflict rate is a first-class metric rather than a guess — strain on the state model is measured, not assumed.
Traces, metrics and logs that join up: Eight custom instruments alongside automatic HTTP, SQL pool, Redis pool and Go runtime telemetry. Every log line carries its trace and span IDs, and a derived field links a log straight through to the trace that produced it.
Migrations that gate the server: Schema migrations run once against a health-checked database and must exit clean before the game server is allowed to start. A half-migrated box cannot serve a match.
Alerting provisioned as code: Alert rules, contact point and notification policy declared in OpenTofu and version-controlled, covering telemetry blindness, memory exhaustion and disk pressure — with CloudWatch alarms held out of band for the case where the in-band path is the thing that died.
Reproducible ARM64 delivery: Images built for Graviton in CI, published to a private registry, and rolled out to the box over SSM — no inbound SSH, no manual step, and log rotation bounded on every container.

Provenance

Where this was made

University of Virginia

School of Engineering and Applied Science · Aug 2021 — May 2024

B.S. Computer Science
GPA 3.88
Rodman Scholar — top 5% of the Engineering class of 2024
AWS Solutions Architect Associate

Instruments

Languages: Java · Go · Python · SQL
Data: PostgreSQL · Redis · Kafka · DynamoDB
Infrastructure: Distributed systems · Microservices · ECS · Lambda · SQS · S3
Practice: REST APIs · Async consumers · CI/CD · Jenkins · Git

Make contact