CrashLoopBackOff from OOMKill: agent diagnoses root cause and patches in 12 seconds. Same incident takes SREs 47 minutes on average. Kube SRE Gym trains small language models to do this autonomously.
Today's infrastructure automation relies on hand-coded runbooks and static rules. They break at scale, can't generalize, and require constant human updates.
Every new failure mode requires a new runbook. Your team spends more time writing playbooks than building product.
Humans diagnose outages through trial-and-error. Average mean time to recovery is 4+ hours for complex infrastructure failures.
Hiring senior SREs is expensive and competitive. You can't hire your way out of infrastructure complexity.
An RL environment where small language models learn to diagnose and fix real Kubernetes production failures — autonomously.
Agents observe cluster state, reason about failures, take corrective actions, and learn from outcomes via GRPO reinforcement learning.
Platformy trains small language models with reinforcement learning in realistic Kubernetes environments, then deploys them to production.
Inject real production failure scenarios into sandboxed K8s clusters: OOM kills, network partitions, misconfigs, resource exhaustion.
Small language models (1-3B params) learn diagnostic reasoning and corrective kubectl actions through Group Relative Policy Optimization.
Claude API acts as a judge, scoring agent responses on accuracy, efficiency, and safety. Only correct fixes earn positive reward.
Trained agents monitor real clusters, detect anomalies, and autonomously execute remediation with human-in-the-loop guardrails.
RL-trained agents are powerful. They're also dangerous if unchecked. Platformy agents are contained at every layer — here's how.
From training RL agents on production incidents to building LLM agent pipelines for executive assessments — here's what I've been building.
Flagship product. Small language models learn to diagnose and repair real K8s production failures through GRPO reinforcement learning. Won 1st place at the Meta Llama Hackathon and presented at the Google Gemini Meetup.
Ongoing research paper on training SRE agents that avoid trap actions — naive fixes that worsen outages. Three contributions: (1) Thompson-sampling harness synthesis (3 auto-synthesized rules achieve 89.7% accuracy, approaching 94.9% hand-written), (2) REx refinement loop with SME feedback (2.8× pass@1 lift on cascade incidents), and (3) deterministic verifiable reward replacing noisy LLM judges.
Designed and built LLM agent pipelines for automated executive assessments across 25+ portfolio companies. Work included RL-based training pipelines, evaluation benchmarks, and real-time inference architecture for delivering AI-generated talent assessments at scale.
The idea came from a 3am page: a CrashLoopBackOff cascading through production, the runbook said "restart the pod," and restarting made it worse. The real fix required reasoning about the system — memory limits, the OOM killer, the deployment spec — not following a playbook. That's exactly what reinforcement learning teaches AI agents to do.
Ashish Ranjan is an SRE and AI engineer. Before Platformy, he built RL-based training pipelines, evaluation benchmarks, and real-time inference systems for LLM agents at a stealth startup (25+ portfolio companies). Kube SRE Gym won the Meta Llama Hackathon and was presented at the Google Gemini Meetup. He's currently writing a research paper on trap-aware SRE agents — agents that refuse to take the naive fix that worsens the outage.
If your on-call rotation is drowning in alerts and your MTTR is measured in hours, let's talk. Book a demo to see RL-trained agents diagnose and fix real cluster failures live.
Pick a time that works for you. All meetings are on Cal.com.
Best for: Quick intro & fit-check.
You'll see a live cluster fail and auto-remediate, ask questions, and decide if Platformy fits your environment.
Best for: SRE leads evaluating the platform.
Full demo: walk through the RL training pipeline, see live incident resolution on a real failure, discuss your use case and integration path.
Best for: Engineering teams going deep.
Dive into GRPO training, harness synthesis, deterministic verifiable rewards, RBAC scoping, and deployment architecture for your cluster.