Lead DevOps Engineer

Job description

Job Title: Lead DevOps Engineer Location: Bengaluru (100% WFO) Job Type: Full-time The problem Skit.ai runs autonomous voice agents for regulated enterprises — India's largest banks and telcos, and US collections operations. Every call is a live distributed system: PSTN/SIP → media server → ASR → LLM → TTS → back, spread across three clouds and multiple vendors, with a conversational response budget measured in hundreds of milliseconds. The platform peaks at roughly \\*\\1 million calls per hour\\\\. Billed minutes grew \\\\5,000x+ in eight months\\\\*. At this scale, infrastructure is not a support function — latency, cost-per-minute, and auditability are product features. When infra degrades, a customer mid-sentence hears silence. We're hiring a Lead DevOps Engineer to own this substrate and keep it ahead of the growth curve. What You'll Own • Multi-cloud substrate: Production infrastructure across AWS, GCP, and Azure. Private interconnects (Direct Connect, Cloud Interconnect, ExpressRoute), transit/hub-spoke topologies, and cross-cloud latency managed as an explicit budget — p95 per hop in tens of milliseconds, not "best effort." • Real-time media plane: Self-hosted LiveKit and SIP infrastructure at scale. Media servers are stateful; you'll design session-affine, event-driven autoscaling (KEDA-class) that survives traffic tripling within an hour. • Model-serving infrastructure: GPU fleets (A100/H100/B200-class) for self-hosted ASR and open-weight LLMs — inference optimization, prefix caching, sticky-session routing, sub-500ms TTFT budgets — alongside managed APIs (Vertex AI/Gemini, Bedrock, Azure). Vendor failover is your design, not your incident. • Reliability & observability: OTel-native tracing (Grafana/Tempo stack), per-turn latency attribution across telephony/ASR/LLM/TTS, automated incident response and self-healing. You'll act as incident commander for infrastructure and write the runbooks you'd want at 3 a.m. • Cost engineering: Cost-per-minute is an SLO here. We cut per-minute serving cost ~18x in six months through caching, rightsizing, autoscaling, and workload re-architecture — you'll own the next 10x. • Security & compliance: Zero Trust across clouds: private endpoints/PrivateLink, IAM/RBAC, secrets management with rotation, WAF/DDoS protection. Operate controls for SOC 2 and ISO/IEC 27001; working command of ISO/IEC 42001:2023 (AI management systems) — hands-on preferred, rigorous theoretical grounding acceptable. You'll face bank and telecom auditors directly, including data-residency requirements. • Technical leadership: Terraform-first IaC standards, production-readiness reviews, mentoring SREs. "Lead" means you raise the floor of the whole team. Problems on our plate right now • Scaling stateful, self-hosted media servers past current concurrency ceilings — HPA on CPU doesn't cut it • Migrating LLM inference from managed APIs to self-hosted open-weight models on GPUs without breaking TTFT budgets • ASR, LLM, and telephony living in different clouds: interconnect topology that keeps the packet path short and private • Multi-region DR that satisfies bank audits without doubling spend If these read as interesting rather than terrifying, keep reading. Must-have • 6+ years hands-on cloud infrastructure; 3+ years operating multiple clouds simultaneously in production; deep expertise in at least two of AWS/GCP/Azure • Real-time audio/video systems in production — WebRTC, SIP/PSTN, or streaming media; you've debugged jitter, not just read about it • Networking depth: VPC/VNet design, load balancing, DNS, NAT; private connectivity (Direct Connect / Cloud Interconnect / ExpressRoute, PrivateLink / Private Service Connect); transit gateways and cross-cloud mesh • Kubernetes at scale (EKS/GKE/AKS), Helm, and scaling \\stateful\\ workloads; service mesh familiarity (Istio/Linkerd) • Infrastructure as Code: Terraform (non-negotiable) across multi-account/multi-project estates; drift is a bug • AI/ML serving in production: GPU allocation and scheduling, inference servers or serverless GPU platforms (vLLM / Triton / Modal / Baseten-class), streaming protocols (WebRTC, WebSocket, gRPC) • Production STT/TTS/LLM API operations: streaming integrations, quota management, multi-vendor failover (Deepgram / Google / Azure / Whisper-class ASR; ElevenLabs / Azure-class TTS) • Security fundamentals: IAM/RBAC, secrets management (Vault or cloud-native), encryption and key rotation • CI/CD: GitHub Actions or GitLab CI with security scanning integrated into the pipeline Strong signal (nice-to-have) • LiveKit, pipecat, Twilio, or comparable real-time platforms; SIP trunking and PSTN integration • KEDA or other event-driven autoscaling used in anger • MLOps: model versioning, canary and A/B rollout • FinOps discipline: reserved/spot strategy, unit-economics reporting • Certifications: AWS SA Professional, GCP Professional Cloud Architect, Azure Solutions Architect Expert • ISO/IEC 42001:2023 exposure What We're NOT Looking For • Single-cloud depth with documentation-level knowledge of the other two • Tool-checklist DevOps without production AI/ML serving scars • "Can learn quickly" as the primary qualification — this role needs day-one production credibility • Anyone who has never traced a packet across a cloud boundary