2,500RPS
Senior Software Engineer · Backend / Platform / Infrastructure
Backend systems, built and operated end to end.
I'm Amit Kumar. I design distributed services, search, and workflow systems—and run the Kubernetes, networking, delivery, and observability layers beneath them.
- Current
- Senior Software Engineer at Coupang
- Focus
- Distributed systems · search · platform reliability
- Operating
- Kubernetes · ArgoCD · Tailscale · Prometheus · Grafana
live platform topologypublic + private ingress
Public trafficinternet-facing services
edgeCloudflareDNS + proxy
routeNGINXTLS ingress
runtimeServicesKubernetes workloads
Private traffictailnet-only applications
clientTailnetWireGuard
routeTailscaleoperator ingress
runtimePrivate appsClusterIP
Service dependenciesapplication state
computeservice
→statePostgres · Redis
Observabilitymetrics path
scrapeworkloads
→storePrometheus
→visualizeGrafana
>90%
lower infrastructure cost
ECS tuning and region migration40%
faster video packaging
Go, gRPC, Redis coordination50%
lower API latency and load
multi-tenant Kong platformSelected systems
From service design to the production path
Each case study focuses on the engineering decisions that survive release: state, latency, isolation, deployment, failure modes, and operating cost.
GitOps control plane
sourceGit
→reconcileArgoCD
→runtimeKubernetes
ingressnetwork policystorageobservability
app of appsself-heal · prune
Infrastructure / GitOps
Personal Kubernetes platform
A GitOps-managed Kubernetes cluster on Oracle Ampere for public services, tailnet-only applications, and development workloads.
- Problem
- Run multiple services with repeatable delivery, private access, persistent state, and enough operational visibility to debug the platform itself.
- Design
- ArgoCD reconciles an app-of-apps tree. NGINX and cert-manager serve public routes, while the Tailscale operator provides ingress for private applications. NetworkPolicy constrains workload traffic.
- Operations
- Immutable image tags, health probes, explicit resources, Prometheus service monitors, provisioned Grafana dashboards, and Git-backed configuration.
- Result
- A fresh cluster can be bootstrapped from one root application and converges to the declared state.
KubernetesArgoCDTailscaleNGINXNetworkPolicyPrometheusGrafana
Live query path
Batmanbest match
Batman Foreverrelated
live endpoint · edit the queryRust API
Search / Backend
Rust typeahead service
A movie-title search service built around a word inverted index and fuzzy candidate ranking.
- Problem
- Return useful matches while a user is still typing, without moving the entire search path into a heavyweight external service.
- Design
- The Rust service loads the title corpus, builds the index, retrieves candidate sets, and ranks fuzzy matches behind a small HTTP API.
- Operations
- The API and frontend are deployed independently with TLS ingress, health probes, resource requests, and immutable container tags.
- Result
- The search path is small enough to run as a low-resource workload on the personal cluster.
RustInverted indexFuzzy matchingDockerHelm
Applied ML / Vector search
Semantic icon discovery
A live icon marketplace with semantic retrieval over generated assets, deployed on the personal cluster.
- Problem
- Keyword matching breaks down when users describe an idea differently from the labels attached to an icon.
- Design
- Sentence Transformers encode natural-language queries with all-MiniLM-L6-v2, and a Pinecone index retrieves the nearest icon candidates for the marketplace UI.
- Operations
- The embedding API runs as a containerized Flask service behind TLS ingress. The model is bundled with the image, while credentials are injected from Kubernetes secrets.
- Result
- Users can discover visually relevant icons by intent instead of exact tags in a working, public demo.
PythonSentence TransformersPineconeVector searchKubernetes
cold-start comparison
containerGo / Knative
~5sisolateRust / V8
<5msstartup costisolation boundary
Compute / Runtime
Container and isolate runtimes
Two serverless execution paths built to understand the boundary between isolation and startup cost.
- Problem
- Reduce cold-start overhead while retaining a useful execution boundary for untrusted or short-lived work.
- Design
- One path uses Go with Knative containers. A second path embeds V8 and uses Rust/WebAssembly-style isolation for smaller units of execution.
- Operations
- The implementations separate runtime lifecycle, request handling, and execution state so startup work can be measured directly.
- Result
- The container path starts in about 5 seconds; the isolate path starts in under 5 milliseconds.
GoRustKnativeV8WebAssembly
multimodal ingestion
ground reportimage · video · text · GPS
semantic pathembeddingsQdrant
geographic pathcoordinatesPostGIS
merge candidatesrelated report clusters
Data / Search
Multimodal media discovery
A newsroom ingestion pipeline for grouping ground reports by semantic and geographic similarity.
- Problem
- Make incoming media discoverable when filenames and manually entered metadata are incomplete or inconsistent.
- Design
- Multimodal models produce embeddings, Qdrant handles vector retrieval, and PostGIS adds geographic constraints before clustering.
- Operations
- The pipeline keeps semantic and location indexes separate, allowing either retrieval path to be inspected and tuned independently.
- Result
- Reporters can find related media by meaning and location instead of relying on exact metadata matches.
PythonMultimodal modelsQdrantPostGISVector search
Personal infrastructure
A cluster I operate, not just describe
The public repository is the source of truth for networking, storage, monitoring, shared services, and application delivery. ArgoCD continuously reconciles it.
cluster traffic and operationsOracle Ampere · Kubernetes
public route
internetClient
→DNS / edgeCloudflare
→TLS routeNGINX ingress
→workloadPublic service
private route
deviceTailnet client
→WireGuardTailnet
→operator proxyTailscale ingress
→ClusterIPPrivate service
continuous deliveryGit → ArgoCD → Kubernetes
service stateApps → PostgreSQL / Redis → OCI volume
observabilityWorkloads → Prometheus → Grafana
Public edge
TLS termination and host-based routing for public services.
- Cloudflare DNS
- NGINX Ingress
- cert-manager
- MetalLB
Private network
Tailnet-only ingress and encrypted routes between cluster services and local machines.
- Tailscale Operator
- Tailscale ingress
- WireGuard routes
- Egress proxies
Control plane
Git-backed reconciliation with automated pruning and self-healing.
- Kubernetes
- ArgoCD
- App of Apps
- Kustomize
- Helm
State and observability
Service-owned state, durable storage, metric collection, and provisioned dashboards.
- PostgreSQL
- Redis
- OCI block volume
- Prometheus
- Grafana
GitOps source of truthgithub.com/AimAmit/infra
Open infrastructure repository Experience
Production systems since 2020
Backend services, search infrastructure, media pipelines, internal platforms, and operational improvements across four engineering teams.
Current role
Senior Software Engineer
Coupang
Dec 2025 — PresentOrder workflows and service reliability
- Contributing to the migration of order cancellation workflows from a legacy Activiti-based framework to Temporal.
- Implemented aspect-based Resilience4j circuit breakers across Spring Boot and WebFlux services to reduce thread exhaustion and cascading failures.
- Built a Slack domain bot using an internal agent framework and MCP to speed up debugging and root-cause analysis.
JavaTemporalSpring BootWebFluxResilience4jMCP
Software Development Engineer 3
Network18
Dec 2023 — Dec 2025Media platforms, search, and internal tooling
- Designed a multimodal ingestion platform using vector embeddings, Qdrant, and PostGIS for semantic and geographic clustering.
- Built a RAG-based VS Code extension backed by a self-hosted model served with vLLM, improving code-assistance efficiency by 25%.
- Built a multi-tenant API platform on Kong that reduced API latency and load by 50% and removed repeated per-service setup work.
- Built an FFmpeg pipeline that converted articles into livestream-ready YouTube video for more than 100 peak concurrent viewers.
PythonQdrantPostGISvLLMKongFFmpeg
Senior Software Engineer
BYJU’S
May 2022 — Dec 2023Video infrastructure and high-throughput services
- Re-architected video packaging in Go with gRPC and Redis-backed multi-master coordination, improving packaging speed by up to 40%.
- Reduced infrastructure cost by more than 90% through ECS tuning and an AWS region migration.
- Built a FIFA quiz service that handled 2,500 requests per second with sub-10ms response time.
GogRPCRedisECSDynamoDBAWS
Software Engineer
Toppr
Jun 2020 — May 2022Search and edge delivery
- Implemented semantic search for academic questions with OpenSearch, BERT embeddings, and approximate nearest-neighbor retrieval at 100ms p95.
- Developed a low-latency typeahead service using optimized Elasticsearch queries.
- Built an image delivery path with S3, CloudFront, Lambda@Edge, and CloudFront Functions for format negotiation and SEO.
OpenSearchElasticsearchBERTS3CloudFrontLambda@Edge
Capabilities
Tools grouped by responsibility
Language and framework choices follow the workload. The recurring concerns are interfaces, state, delivery, networking, and observability.
Services
GoPythonRustJavaSpring BootFastAPIDjango
Interfaces
RESTgRPCGraphQLWebSocketMCPWebAssembly
Data and search
PostgreSQLRedisDynamoDBCassandraElasticsearchOpenSearchQdrant
Infrastructure
KubernetesDockerArgoCDAWSGCPCloudflareNGINXCI/CD
Networking
TailscaleWireGuardDNSIngressCDNLoad balancingNetworkPolicy
Selected writing
Field notes from systems I run
Practical notes on infrastructure and development workflows, written from implementation rather than abstraction.