$whoami — engineering journal & portfolio
I build the infrastructure behind document intelligence — high-performance backend systems, cloud-native platforms, and production LLM serving that stays reliable under real load.
I like making infrastructure boring: predictable deploys, honest dashboards, and GPUs that earn their cost.
Backend Engineer @ DHL Global Forwarding · Product Development Team Leader @ AIDocBuilder
01
High-throughput services and PostgreSQL performance — systems that hold up under real load.
Python · Django · NestJS · PostgreSQL
02
Kubernetes on AKS, OpenShift, and self-managed RKE2 — GitOps with Argo CD, CI on Tekton and GitHub Actions.
AKS · OpenShift · RKE2 · Argo CD
03
Production vLLM serving, GPU resource management, KV-cache and batching optimization.
vLLM · GPU scheduling · KV-cache
4+ yrs
production backend
28
QA namespaces on self-managed K8s
20+
services through in-cluster CI
DEV→PROD
enterprise release pipeline owned
01 /
Four-plus years of backend work, told without the buzzwords.
I'm a backend engineer at DHL Global Forwarding, contracted through AIDocBuilder Inc., where I also lead the product development team. My work sits where enterprise logistics meets document intelligence: systems that move real freight paperwork through extraction pipelines, and the infrastructure those systems run on.
I started in classic backend engineering — Python, Django, Node.js, and a lot of PostgreSQL. The systems grew, and so did the failure modes, which pulled me toward the platform underneath: Kubernetes on AKS, OpenShift, and self-managed RKE2, GitOps delivery with Argo CD, and CI pipelines — GitHub Actions and in-cluster Tekton — that teams can actually trust.
The last stretch of that path has been AI infrastructure. I run production LLM inference with vLLM — GPU node pools, KV-cache tuning, batching behavior, cost-per-document math. It's the same discipline as database performance work, just with more expensive hardware.
This site is my engineering journal. The war stories are real production incidents and what they taught me; the case studies and writing are the longer-form version of the same habit — learning in public, with specifics.
worked with: DHL Global Forwarding · AIDocBuilder · Azure (AKS) · Red Hat OpenShift · NVIDIA GPU infrastructure
02 /
From backend services to the platform underneath them. Select an entry to expand it.
Built and maintained REST APIs backed by PostgreSQL. Learned the fundamentals the only way that sticks: being on the hook when queries got slow and deploys went sideways. Picked up Docker and the habit of reading query plans before blaming the database.
Python · Django · PostgreSQL · REST APIs
Designed ingestion and extraction services in Node.js/NestJS alongside existing Python systems. Owned the data layer: schema design, indexing strategy, and the queue-based architecture that let processing scale horizontally instead of vertically.
Node.js · NestJS · RabbitMQ · Schema design
Moved into enterprise scale: systems where a slow endpoint delays real shipments. Focused on database performance under production load, API reliability, and integrating document intelligence into logistics workflows that operations teams depend on daily.
Enterprise systems · Database performance · Logistics domain
Ran workloads on AKS and OpenShift, moved deployments to GitOps with Argo CD, and rebuilt CI/CD on GitHub Actions. The shift from 'my service' to 'the platform every service runs on' changed how I think about reliability, cost, and developer experience.
Kubernetes · AKS · OpenShift · Argo CD · GitHub Actions
Took on team leadership for AIDocBuilder's product engineering while owning production LLM serving: vLLM on GPU node pools, KV-cache and batching tuning, and the unglamorous work of making inference costs predictable. Leading people and running GPUs turn out to need the same skill — capacity planning.
Team leadership · vLLM · GPU scheduling · LLM serving
Shipped the document-AI platform into a locked-down enterprise network across DEV, UAT, and PROD — where every network policy surfaces as an application bug. Migrated CI/CD from Bitbucket to GitHub without a maintenance window, and ran the QA cluster's reliability campaign, turning recurring incidents into limits, garbage collection, and watchdogs. The direction from here: platforms that make AI deployment boring.
AKS · RKE2 · Longhorn · GitOps · Incident automation
03 /
Real production incidents: what broke, how I found it, and what it taught me.
Backend views started throwing NoneType errors on Redis reads, and the Redis pod's logs were spotless — because OOMKill is a SIGKILL, and dead processes don't write farewell notes.
Redis · Kubernetes · Django—3 min read
Zero snapshot-cleanup jobs plus 200% over-provisioning let dead Longhorn snapshots pin ~100GB of ghost blocks under the Harbor registry volume. It filled a 196GB disk, the filesystem went read-only, and every build in the cluster started failing with I/O errors.
Longhorn · Harbor · Kaniko · ext4—4 min read
Seven workers hit disk pressure, and every standard fix made it worse: restarted agents flapped back NotReady in minutes, kubelets hung silently for 30+ minutes on startup, and Harbor — the registry every recovering pod needed — went down with the cluster it runs on.
RKE2 · Longhorn · Harbor · Argo CD · kubelet—4 min read
04 /
Deep dives into systems I've designed and shipped — architecture, trade-offs, results.
Deploying a document-intelligence platform onto customer-managed AKS behind an enterprise firewall that never says no — it silently drops traffic. Every network policy surfaced as an application bug; the work was proving which layer was lying.
AKS · Azure · Python · Django · SMTP · Artifactory
A rapid-QA tier of 28 namespaces on a self-managed cluster where Longhorn shared the root disk with kubelet and reserved nothing. One full disk became a cluster-wide eviction cascade; this is the reliability campaign that followed.
RKE2 · Longhorn · Argo CD · Tekton · Harbor · KEDA
05 /
Notes on backend engineering, Kubernetes, and AI infrastructure.
A CORS failure that was a file permission. A registry auth error that was a full disk. In layered systems the error tells you where the failure surfaced, not where it lives — four production examples and the habit that cuts through them.
OOMKill is a SIGKILL: the process gets no chance to log, so the pod's logs look innocent while the whole fleet throws NoneType errors. Where the evidence actually lives, and the write-time habits that prevent the class.
Longhorn said the volume was healthy. The pod got EIO on every write. Both were right — about different layers. What control-plane health actually attests, and the e2fsck recovery that preserves data.
06 /
Organized by domain, with the context each tool was actually used in.
Services that stay fast when the data grows.
Python / Django
APIs and document pipelines since 2021
Node.js / NestJS
document-processing services at AIDocBuilder
PostgreSQL
schema design, query plans, connection pooling under load
Redis
job state and caching — TTL discipline learned in production
RabbitMQ
work queues: ack semantics and poison-message handling
The paved road every service runs on.
Kubernetes
production workloads on AKS, OpenShift, and self-managed RKE2
AKS
enterprise deployments inside locked-down customer networks
OpenShift
enterprise clusters — first responder for infrastructure issues
Argo CD
GitOps delivery for 28 applications, repo migrations included
Tekton / Harbor
in-cluster CI and registry for 20+ services
Longhorn
distributed storage operations: snapshots, replicas, salvage
GitHub Actions
CI/CD from build through release promotion
Docker
multi-stage builds, image size and cache discipline
Making LLM inference predictable — in latency and in cost.
vLLM
production inference serving for document AI
LLM serving
self-hosted extraction models on GPU infrastructure
GPU optimization
maximizing utilization of existing GPU capacity
KV-cache tuning
cache and memory budgeting under real workloads
07 /
I'm open to conversations about backend and platform engineering roles, AI infrastructure work, and hard production problems. The fastest way to reach me is email — I read everything.