AI Platform Engineer · Bengaluru, India

SAIKUMAR

Inference wired. Latency obsessed. Fluent in the strange dialect of GPUs, YAML and token budgets. I build the platforms that get AI from demo to production for airlines, healthcare and payments.

Try my gateway → LinkedIn GitHub
SK
PASSENGERSAIKUMAR
FROMDEVOPS
TOAI PLATFORM
GATEOPEN
About

SaiKumar

say it: sigh · KOO · mar

Senior Platform Engineer for AI infrastructure at Delta Air Lines. Before that, MLOps and multi-cloud Kubernetes at Carelon Global Solutions in healthcare, and SRE at NexInfo.

But what do I really do?

I build the parts of AI nobody demos: the gateways, quotas, rollouts, retrieval pipelines and dashboards that let models ship safely and stay up.

What I bring: reliability → in serving, in scaling, in spend.

Delta Air LinesCarelon Global SolutionsNexInfo Solutions
Measured impact

Numbers from production

42%

Lower document-intelligence latency from a RAG pipeline on AWS Bedrock with MCP servers

Delta Air Lines
500+

Production services moving to a multi-tenant OpenShift platform

Delta Air Lines
30%

Lower compute spend from that platform migration

Delta Air Lines
5D→1H

Service onboarding time after GitOps Day-0/Day-1 automation

Delta Air Lines
99.95%

Uptime for Kafka and Spark on Kubernetes at 50% higher throughput

Carelon Global Solutions
35%

Lower MTTR after standardizing on Splunk and OpenTelemetry

Carelon Global Solutions
How I build

Four worlds, one platform

Airlines, hospitals, payment networks and GPU clusters all need the same thing from AI: fast, safe and affordable. Scroll through what that looks like.

01 · Airline

Answers at airline scale

At Delta, a RAG pipeline on AWS Bedrock with custom MCP servers made document intelligence 42% faster, on a platform that 500+ services run on.

02 · Healthcare

Sensitive data, kept in bounds

At Carelon, Kubeflow MLOps and multi-cloud Kubernetes kept data pipelines at 99.95% uptime, with identity-based access through Teleport.

03 · Payments

Agents that recommend, people who decide

My open-source fraud platform: an agent gathers evidence over MCP, guardrails strip card numbers and block prompt injection, and every case goes to an analyst.

04 · Platform

GPUs, gateways and budgets

Under all of it: vLLM on Kubernetes, a Go gateway that routes, caches and meters every token, and SLO alerts before anyone notices a slowdown.

Flight log

Seven years, one heading

From keeping production up, to running platforms, to building the infrastructure AI runs on. Scroll to fly the route.

NXI2019 CRL2023 DAL2024 NEXTAI PLATFORM
2019 – 2022NexInfo SolutionsDevOps SRE: CI/CD, observability, GKE
2023 – 2024Carelon Global SolutionsDevOps: Kubeflow MLOps, EKS and AKS
2024 – nowDelta Air LinesSenior Platform Engineer, AI infrastructure
BoardingYour team?AI platform, LLM inference, ML infrastructure
Playground

Send a prompt through my gateway

This runs the same routing rules as the Go AI gateway in my payments project: auth, rate limit, token budget, complexity routing and cache. Send the same prompt twice to see a cache hit.

Request

600 / 600 requests left this minute

Gateway trace

Authwaiting
Rate limitwaiting
Budgetwaiting
Routerwaiting
Cachewaiting
Upstreamwaiting
MODEL—
TOKENS IN—
COST—

Simulated in your browser; no model is called. Prices match the project's example config.

Case studies

Airline. Healthcare.

Two industries where downtime is expensive and data is sensitive. Here is what I built in each.

AirlineDelta Air Lines · Dec 2024 – now

GenAI infrastructure for a global airline

ROLESr Platform Eng
FOCUSAI infra
SERVICES500+
42%lower RAG latency
100+teams on the platform

What I built

  • A high-throughput enterprise RAG pipeline on AWS Bedrock, with custom MCP servers for secure agentic data retrieval.
  • Migration of 500+ production services to a multi-tenant OpenShift (ROSA) platform, cutting compute spend by 30%.
  • Self-service developer platform for provisioning, secrets and deployments, used by 100+ teams.
  • GitOps with GitLab CI and Helm that took onboarding from 5 days to under 1 hour.
  • OpenTelemetry pipelines for 40% faster root-cause analysis, plus progressive delivery and auto-remediation.
AWS BedrockMCPOpenShift ROSAGitLab CIHelmOpenTelemetryAWS CDK
HealthcareCarelon Global Solutions · 2023 – 2024

MLOps and multi-cloud for healthcare

99.95%data platform uptime
50%faster release cycles

What I built

  • Early MLOps adoption with Kubeflow, plus Crossplane abstractions across AWS and Azure.
  • Large-scale microservices on EKS and AKS with ArgoCD GitOps, for 50% faster releases.
  • Kafka and Spark on Kubernetes at 99.95% uptime with 50% more throughput.
  • KEDA event-driven autoscaling and Karpenter self-healing nodes, cutting compute waste by 20%.
  • Python access automation that removed 80% of manual effort for 500+ engineers, secured with Teleport.
KubeflowCrossplaneEKSAKSArgoCDKEDAKarpenterTeleport
Open-source builds

AI platform projects

End-to-end builds, each with serving, a gateway, agents, guardrails and evals. Each card shows its real status.

PaymentsCode ready · GPU benchmarks pending

Agentic fraud-investigation platform

  • AI gateway in Go: complexity routing, per-team quotas and token budgets, caching, SSE streaming, failover, cost per tenant
  • Agent gathers evidence through an MCP server and drafts cases for analyst approval; it never moves money
  • Guardrails for PII, prompt injection and output schema, with an eval gate in CI; vLLM on KServe with canary rollouts
Go · Python · KServe · vLLM · MCPRepository →
AirlinePlanned

Flight-disruption recovery agent

  • MCP servers over schedule, crew and weather data built from public US DOT on-time records
  • Agent proposes rebooking and crew-swap options; an ops controller approves
  • Served through the Go gateway with p95 latency SLOs during disruption spikes
LangGraph · MCP · KServe · KEDA
HealthcarePlanned

PHI-safe clinical RAG platform

  • RAG over synthetic patient records from Synthea, exposed through a FHIR MCP server
  • PHI redaction guardrails, audit logging of every model call, faithfulness evals
  • Tenant isolation on Kubernetes with Vault-managed secrets
FHIR · MCP · Ragas · Vault
RetailPlanned

Shopping-assistant RAG platform

  • Hybrid search and reranking over a product catalog with live price and stock tools
  • Embedding experiments in MLflow, promoted through a Ragas gate
Qdrant · MLflow · KServe
NetworkingPlanned

AgenticOps troubleshooting agent

  • MCP servers over telemetry, syslog and topology from a containerlab network
  • Kubernetes operator in Go; GPU sharing with MIG and time-slicing
Go · MCP · GPU Operator
Toolkit

What I work with

LLM servingvLLM, KServe, continuous batching, prefix caching, SSE streaming
AI gatewaysRouting, rate limits, quotas, token budgets, cost tracking
RAG and agentsAWS Bedrock, MCP, LangGraph, vector search, guardrails, evals
KubernetesOpenShift, EKS, AKS, GKE, Karpenter, KEDA, Istio, Crossplane
MLOpsKubeflow, model rollouts, CI/CD for models and prompts
DeliveryGitLab CI, ArgoCD, Azure DevOps, Terraform, AWS CDK
ReliabilityOpenTelemetry, Prometheus, Grafana, Splunk, SLOs
LanguagesGo, Python
Work with me

Let's get your models to production.

Open to AI platform, LLM inference and ML infrastructure roles. AWS Certified Solutions Architect – Professional.

LinkedIn → GitHub →