Selected work

Shipped, measured, still running.

Four systems, four industries, one pattern: senior team in, hard problems out, production metrics shared with you.

01 / Fintech
order-gateway · prod
p99 latency42 ms
throughput18.4k/min
uptime 30d99.99%

Liquid Edge — real-time trading desk

An open broker platform was melting its message queue in every price spike. We rebuilt the order path on Go + Kafka with an idempotent gateway, deployed behind a canary pipeline, and moved from "urgent incident" to "boring Monday".

GoKafkaKubernetesClickHouseTerraform
4× peak volume99.99% uptime0 downtime deploys42ms p99 events
Read case study
02 / AI
copilot-api · eval
answer median1.9 s
hallucination1.6%
auto-resolved71%

Support Copilot — RAG over 40k docs

A B2B SaaS wanted an AI support agent, not a demo. We built a hybrid retrieval pipeline over their knowledge base with a labelled eval set, streaming answers, citations and a guardrail layer that routes confident failures to humans.

PythonLangChainpgvectorFastAPIOpenAI
8× faster answers<2% hallucination40k docs indexed
Read case study
03 / Gaming
lobby-svc · prod
concurrent104k
p95 match28 ms
incidents0

Skybound Arena — match backend that survived launch

Launch day brought 6× expected concurrency. The matchmaking, presence and economy services we built absorbed the spike without a single incident ticket — then the telemetry pipeline told product exactly where players quit.

GoRedisKafkaAWSClickHouse
100k+ concurrent0 launch incidents28ms p95 matchmaking
Read case study
04 / Data
metrics-lake · prod
freshness5 s
dashboards17 → 1
teams4 / 4

Meridian — one metric lake for a logistics company

Seventeen dashboards disagreed about revenue. We built real-time ingestion, a dbt-modelled warehouse and a single metric layer, then migrated every team onto one definition — with freshness tests that page someone when data goes stale.

KafkadbtSnowflakeAirflowMetabase
17→1 sources of truth5s event freshness4 teams migrated
Read case study

Want case studies with your name on them?

Every project ships with a shared war-room, public metrics and a reference you can call. That's the model.

Start yours Read engineering notes