Rahul Garg
I build RL environments and evals that train AI agents
Dockerized environments · verifiable graders · DevOps-domain tasks for frontier-model post-training
Senior infrastructure engineer (SDE-3) turned agentic-task specialist. Currently the reviewer gate for agentic DevOps tasks and evaluation rubrics at a US-based AI lab (under NDA) — promoted from task creator to QC reviewer for frontier-model training data. Previously built Tessell's cloud-native DBaaS platform on GCP from scratch.
The distributed-systems depth is what makes the environments realistic: five years of Kubernetes, CI/CD, and cloud debugging means my tasks reflect how production infrastructure actually fails. Available for contracts via Ananta Systems.
◆ Work with me
Per-environment builds
Fixed-scope Dockerized RL environments with task specs, hidden verifiers, and difficulty tiers — delivered ready to train against.
Eval design & QA review
Design of eval rubrics and graders, plus rigor review of existing environment suites: reward-hacking audits, verifier hardening, difficulty calibration.
Agentic-task development
Hourly development of agentic tasks in the DevOps / infrastructure domain: CI/CD, Kubernetes, Terraform, cloud debugging.
◆ Skills
◆ Experience
Promoted from task creator to reviewer (Jul 2026): QC-reviewing agentic DevOps tasks, Dockerized RL environments, and eval rubrics produced by a 100+ contractor pool. Previously designed those tasks myself — recognized as a top performer.
Built the cloud-native Database-as-a-Service platform on GCP from scratch; redesigned the control plane, cutting infrastructure management overhead by 90%.
Built a Blockchain platform using Hyperledger Fabric, deployed on Kubernetes.
◆ Featured Projects
View all →DevOps Gym
RL EnvironmentsSuite of Dockerized, verifier-graded DevOps tasks for training and evaluating AI agents — broken CI pipelines, failing K8s deployments, crashing services — each with a hidden test-based grader and difficulty tier.
K8s Debug Evals
EvalsReproducible eval harness for measuring frontier models on Kubernetes debugging tasks: pass@1 scoring, a failure taxonomy, and verifiers designed to resist reward hacking.
K8s Job Scheduler
Distributed SystemsGo-based HTTP API server for prioritizing and submitting Kubernetes jobs — the orchestration depth behind my environment work: priority queues, concurrency control, client-go.
HybridRAG
LLM SystemsCost-optimized enterprise search engine with three-layer architecture — evidence of LLM-systems work: ingestion pipeline, hybrid retrieval, and cost-aware model routing.