AI Film Studio Backend Setup

I am building ROLLCALL, an AI film/video-generation platform. The website/frontend already exists.

I need a senior GPU/ML infrastructure engineer to build a clean production V2 backend for Wan2.2 I2V A14B. This is not a website-design project.

We have already proven that Wan2.2 A14B can generate usable video on an NVIDIA A100 80GB. The previous prototype used RunPod, Python, PyTorch/CUDA, FastAPI/Uvicorn and FFmpeg, but I do not want to continue patching a fragile experimental environment.

The engineer may recommend RunPod, Lambda Cloud, CoreWeave, AWS/GCP/Azure or another appropriate GPU provider, but must justify the choice specifically for a large Wan2.2 A14B workload.

Required architecture:

ROLLCALL Website → API → Persistent Job Queue/Database → Disposable GPU Worker → Wan2.2 I2V A14B → FFmpeg/QC → Object/Persistent Storage → ROLLCALL

Requirements:

NVIDIA A100/H100-class production inference
Wan2.2 I2V A14B
Python/PyTorch/CUDA
reproducible Docker-based deployment
pinned/compatible dependencies
persistent model storage
model loaded once and kept resident between jobs where appropriate
FastAPI or equivalent production API
immediate job ID creation
queued/generating/processing/completed/failed states
persistent job/segment state outside disposable GPU compute
idempotent jobs/retries
GPU crash/OOM recovery
automatic restart/recovery
FFmpeg + ffprobe output validation
persistent/object storage for generated media
QC frames and production manifests
replacement GPU must resume incomplete work without regenerating completed segments
monitoring/logging
GPU usage/cost monitoring
automatic shutdown/scaling strategy
complete integration with the existing ROLLCALL website

The GPU worker must be disposable. If the GPU instance disappears completely, production state and completed work must survive.

First acceptance milestone:

ROLLCALL Website → Job Created → GPU Worker → Wan2.2 I2V → Real MP4 Generated → FFmpeg/QC Validation → Persistent Storage → Job Completed → Video Plays on ROLLCALL.

I want this first end-to-end workflow operating within the first 1–2 working days, followed by production hardening.

Final acceptance requires:

Real end-to-end video generation.
Worker restart/replacement recovery test.
No loss or duplication of completed segments.
Reproducible deployment from documentation/configuration.
Exact Python/PyTorch/CUDA/package versions documented.
ROLLCALL website successfully plays the generated output.

Screening question — start your proposal with ROLLCALL-V2:

Explain how you would make a Wan2.2 A14B GPU worker disposable while keeping the model warm during active operation. Explain where models, queue/job state, source images, segments and final MP4s would live, and exactly what happens if the GPU dies after completing 7 of 24 segments.

Also tell me:

which GPU/cloud provider you recommend and why;
A100 vs H100 recommendation;
expected implementation time;
fixed-price estimate;
and examples of large diffusion/video inference systems you have personally deployed.

I am looking for completion in days, not weeks.

Do not apply with a generic AI/cloud proposal.