SEEKING WORK | India, remote | Part-time | Overlap with US, UK, EU hours
I build on-prem GPU clusters and private AI apps that run entirely on the client's own hardware, so their data never leaves the building. About 10 years full-stack, the last few on GPU infrastructure.
Currently running a live production cluster: 700-800 users sharing a small physical GPU pool. GPU-backed JupyterHub, single sign-on, per-user quotas, batch scheduling, private LLM chat and RAG system with LLM Evals and LLM Observability.
What I do:
1. GPU clusters on bare metal - Kubernetes across bare servers, full NVIDIA stack: GPU Operetor, device pluugin, container toolkit, DCGM - GPU sharing: time-slicing, MIG, per-user quotas, idle-GPU reclaim - Scheduling: Kueue, so multi-day batch jobs and short interactive sessions share the same pool - Multi-tenancy: Keycloak SSO, namespace isolation, resource quotas - Operations: Prometheus and Grafana, self-healing, unattended restart after power loss
- In progress: Ceph, so one pool of drives serves S3, file and block
2. Private AI apps, on your own hardware - LLM chat: Ollama, Open WebUI - Image and video generation: ComfyUI, Flux, Wan 2.1, Stable Diffusion - Agents: LangGraph, LangChain - Document RAG search, written from scratch: multi-tenant libraries,per-user access control, cited answers, model stays on-prem
3. Data platforms - Pipelines that turn large, messy, real-world data into somethingqueryable - Recent example: government land records into clean data products for public-sector clients - Full-stack web portals and dashboards on top (TypeScript, Python, React, Node)
Email: [email protected]
GitHub: https://github.com/tapudp
LinkedIn: https://linkedin.com/in/pbdivyesh
I build on-prem GPU clusters and private AI apps that run entirely on the client's own hardware, so their data never leaves the building. About 10 years full-stack, the last few on GPU infrastructure.
Currently running a live production cluster: 700-800 users sharing a small physical GPU pool. GPU-backed JupyterHub, single sign-on, per-user quotas, batch scheduling, private LLM chat and RAG system with LLM Evals and LLM Observability.
What I do:
1. GPU clusters on bare metal - Kubernetes across bare servers, full NVIDIA stack: GPU Operetor, device pluugin, container toolkit, DCGM - GPU sharing: time-slicing, MIG, per-user quotas, idle-GPU reclaim - Scheduling: Kueue, so multi-day batch jobs and short interactive sessions share the same pool - Multi-tenancy: Keycloak SSO, namespace isolation, resource quotas - Operations: Prometheus and Grafana, self-healing, unattended restart after power loss
- In progress: Ceph, so one pool of drives serves S3, file and block
2. Private AI apps, on your own hardware - LLM chat: Ollama, Open WebUI - Image and video generation: ComfyUI, Flux, Wan 2.1, Stable Diffusion - Agents: LangGraph, LangChain - Document RAG search, written from scratch: multi-tenant libraries,per-user access control, cited answers, model stays on-prem
3. Data platforms - Pipelines that turn large, messy, real-world data into somethingqueryable - Recent example: government land records into clean data products for public-sector clients - Full-stack web portals and dashboards on top (TypeScript, Python, React, Node)
Email: [email protected]
GitHub: https://github.com/tapudp
LinkedIn: https://linkedin.com/in/pbdivyesh