AI Infrastructure Engineer
6 projects across AI infrastructure, OpenStack private cloud, and VoIP telephony —
from GPU clusters and LLM pipelines to SIP agents and production SLOs.
| 01 | Voice Agentic AI Platform | Real-time voice assistant on Kubernetes — Browser mic → Whisper STT → Mistral 7B ReAct agent (weather, search, calc tools) → Edge TTS. GitOps with ArgoCD + vLLM GPU serving. | KubernetesArgoCDvLLM |
| 02 | LLM Fine-Tuning Platform | Self-service LoRA fine-tuning — upload dataset, get model auto-evaluated and promoted to MLflow. 5-step Argo Workflow DAG: ingest → preprocess → KubeRay train → eval → register. | KubeRayMLflowArgo |
| 03 | AI Platform Ops | Production operations layer for Projects 1 & 2 — custom kopf operator auto-deploys MLflow models, KEDA scales Ray workers to zero, Argo Rollouts canary gates, OTel traces, SLO burn-rate alerts. | kopfKEDAOTel |
| 04 | HA OpenStack Deployment | Production private cloud with Kolla-Ansible — 3-node HA controller cluster (Keepalived VIP + HAProxy + Galera + RabbitMQ), 4 compute nodes, 3-node Ceph storage. Automated deploy & health scripts. | OpenStackCephKolla |
| 05 | GPU Cloud for AI/ML | GPU-enabled OpenStack — NVIDIA A100/T4 PCI passthrough via vfio-pci, Ironic bare-metal provisioning, VLAN 200 training network (MTU 9000 for NCCL), Heat templates for self-service CUDA instances. | IronicVFIOHeat |
| 06 | AI Phone Agent | SIP phone calls handled by an AI agent — FreeSWITCH mod_audio_stream pipes RTP to Python bridge in real time, WebRTC VAD detects speech end, same Whisper→Mistral→TTS pipeline answers the call. | FreeSWITCHSIPVAD |
Real-time voice assistant on Kubernetes. Browser microphone to audio response in under 4 seconds — Whisper transcribes speech, Mistral reasons and calls tools, Edge TTS speaks the answer back.
Self-service training platform — upload a dataset, get a fine-tuned model auto-evaluated and registered in MLflow. A 5-step Argo Workflow DAG handles preprocessing, distributed training, evaluation, and promotion.
Closes all production gaps in Projects 1 and 2. Five components that separate a demo from a system that runs at 3 AM without anyone watching.
| Custom K8s Operator | MLflow → K8s auto-deploy in 30s via kopf @kopf.daemon |
| OTel + Tempo | STT / LLM / TTS waterfall traces — blame the right service in 30s |
| KEDA Scale-to-Zero | Ray workers idle 16h/day → ~60% GPU cost savings |
| Canary Rollouts | 10% → 50% → 100% with Prometheus analysis gates, auto-abort |
| SLO Alerts | Google SRE burn rate formula — fast burn 14×/1h, slow burn 6×/6h |
Full private cloud deployment: 3-node HA controller cluster (Keepalived VIP + HAProxy + MariaDB Galera + RabbitMQ), 4 compute nodes, 3-node Ceph storage backend. Automated deploy scripts with pre-flight checks and post-deploy resource provisioning.
Extends OpenStack for self-service GPU compute. NVIDIA A100/T4 PCI passthrough via vfio-pci, Ironic bare-metal for workloads that can't tolerate hypervisor overhead, VLAN 200 training network with MTU 9000 for NCCL distributed training, and Heat templates for one-command GPU instance provisioning with CUDA + PyTorch.
| PCI Passthrough | NVIDIA A100 & T4 via vfio-pci — full GPU memory, no hypervisor overhead |
| Ironic Bare Metal | iPXE boot provisioning for large-model training nodes |
| NCCL Training Net | VLAN 200, MTU 9000 jumbo frames — optimised for AllReduce collective ops |
| Heat Templates | One command: GPU instance + CUDA 12.4 + PyTorch cloud-init |
Bridges a SIP telephone network to the Kubernetes AI voice pipeline. A caller dials in, FreeSWITCH handles SIP/RTP, and mod_audio_stream streams audio in real time to a Python bridge — through the same Whisper STT → Mistral ReAct → TTS pipeline. The AI answers phone calls.
// open to AI infrastructure roles · Ahmedabad, India