01 — Start with data that doesn't lie.
Postgres with read replicas and indexing strategy. Redis out front as a distributed cache. Baseline: 60% fewer DB queries, 50% faster APIs, reads 60% snappier.
I build AI agentsyou can actually trustin production.
Software engineer with a decade of curiosity and a taste for systems that refuse to fall over. I spend my days making agentic software useful, measurable, and reliable — not just impressive in a demo.
Scroll, slowly →↓
§ 02Philosophy
The gap between a flashy demo and a production system is measured in observability, evaluations, and a thousand small decisions nobody tweets about.
Agents are products.
They deserve SLOs, regression suites, and a paging rotation — same as any other service with users.
Context is the feature.
Retrieval strategy, conversation state, and prompt shape determine whether a user trusts you twice.
Numbers over vibes.
Offline evals gate the deploy. Online evals gate the rollout. Vibes are for the launch post.
Boring where it counts.
Postgres, Kafka, Redis, a circuit breaker. The exciting part sits on top of a foundation that never surprises you.
§ 03The stack, built as you scroll
05Agent layerplanning · tools · summaries
04Retrievalsemantic chunking · citations
03Gateway & servicesSpring Boot · Gin · Kubernetes
02Event busKafka · OAuth 2.0
01Data planePostgres · Redis · replicas
02 — Make services talk without shouting.
Kafka for event-driven flow. +30% throughput, −10% inter-service latency. OAuth 2.0 for 5M+ daily logins with 20% fewer failures.
03 — A gateway that survives the weekend.
High-throughput microservices in Spring Boot + Gin on Kubernetes. 2M+ concurrent users. 99.99% uptime. Legacy translation for 50M requests/day at 25% lower cost.
04 — Retrieval you can argue with.
LlamaIndex over internal tools, code, and knowledge bases. Ingestion, semantic chunking, embeddings, vector search. Answers arrive with citations — or they don't arrive at all.
05 — Agents that know when to stop.
Prototyped in Dify, hardened with ADK. Specialized agents plan steps, call tools, and aggregate. ~70% less time finding internal info. 17% fewer tokens. Engineers & PMs actually use it.
§ 04Where I've been useful
2023 — present · San Francisco Bay Area
Software Engineer, 2K Games
- Shipped an internal AI Slack assistant — agentic RAG over tools, code, and knowledge. −70% time to find info, −17% token cost per conversation.
- Designed & shipped a commerce microservice at 2M+ CCU with 99.99% uptime on Spring Boot + Gin + K8s.
- Tuned the data plane — HikariCP, PG read replicas, Redis cache: −35% DB latency, −60% queries, 20K+ TPS.
- Built a gateway translation layer handling 50M req/day, phased out legacy services, −25% cost.
- Reliability work — Resilience4j circuit breakers, Kafka event bus (+30% throughput), OAuth 2.0 for 5M daily logins.
- Leading beyond my ticket — owning epics, writing tech designs, and coordinating contractors on delivery.
§ 05Signature work
Project 01 · Internal Platform
The Slack assistant that actually reads the docs.
An agentic RAG assistant over our private tools, code, and knowledge. Engineers ask, answers come back with citations — and the bill stays reasonable.
Input
Slack@ai — natural language
Conversation statemulti-turn context window
Reasoning
Planner agentADK · decides which tools to call
Docs agentConfluence · Notion
Code agentrepo search · PR lookup
Ops agentincident triage
Retrieval
Semantic chunkerLlamaIndex
Vector storeembeddings · similarity
Citation weaverevery claim, sourced
Guardrails
Context optimizer−17% tokens
Observabilitytraces · token ledger
Eval harnessoffline + online
−70%time finding internal info
−17%token cost per conversation
Eng + PMdaily-active adoption
0hallucinated citations in prod*
* caught by the eval harness before rollout. That's the whole point.
Project 02 · Platform Infrastructure
A commerce platform that doesn't flinch at launch day.
High-throughput microservices powering in-game commerce for 2M+ concurrent players with four nines of uptime. Boring on purpose.
Edge
API Gateway50M req / day
OAuth 2.05M daily logins
Resilience4jcircuit breakers
Services
Commerce serviceSpring Boot · Gin
Kubernetesauto-scaling
Kafkaevent-driven
Data
PostgreSQLread replicas · indexes
Redisdistributed cache
HikariCPconnection pool
Ops
Jenkins · CI/CDauto deploys
MonitoringSLOs · alerting
Cost controls−25% gateway
2M+concurrent users
99.99%uptime
20K+ TPSquery capacity
−50%API response time
§ 06Making agents honest · side work
Harness engineering, observability, and evaluations — the parts nobody films a demo about.
Outside of 2K, I've been working with a Gen-AI startup on the infrastructure that turns an agent from a party trick into a product:
- Harness engineering — the scaffolding an agent runs inside: tool registries, retries, budgets, timeouts, replay.
- Observability — tracing every step, token, tool call and latency hop so failures are legible.
- Offline evaluations — regression suites with curated cases; a PR doesn't merge if the scoreboard drops.
- Online evaluations — shadow traffic, LLM-as-judge, user-signal telemetry, all feeding the next dataset.
eval-suite/mainPASS
| faithfulness | 0.92 | |
| citation_recall | 0.88 | |
| task_completion | 0.81 | |
| tool_accuracy | 0.95 | |
| p95_latency | 1.8s | |
| cost/query | $0.014 |
+ scenarios regenerate on every commit · drift gates the merge
If an agent can't be measured, it can't be shipped.
If it can't be shipped, it isn't real.
§ 07A quiet ledger of impact
2M+CCU supported
99.99%uptime on commerce
50Mgateway req / day
20K+TPS
5Mdaily OAuth logins
−70%time to find info
−60%DB queries
−50%API latency
−35%DB latency
−25%gateway cost
−17%token cost
1000×pipeline speedup (GU)
§ 08Arrangements, after hours
Engineering and arranging share the same bones — voice-leading is just dependency resolution in a nicer font. A few pieces I've been carrying around.
01
02
03
04
05
06
07
video2:09(Hit ▶. R2 ships the bytes; the rest was shipped after hours.)
§ 09Education
Georgetown UniversityM.S. Computer Science · GPA 3.96 · 2021–2023
Beijing Jiaotong UniversityB.Eng. Computer Science & Technology · 2016–2020
§ 10Let's talk