AI Engineer, Agentic Ad Creative (Multimodal)

NewsBreak•Mountain View, CA

10h•$120,000 - $220,000

About The Position

NewsBreak is seeking an AI Engineer to own and rebuild their creative pipeline, transforming it into an agentic system. This role involves developing an agent that generates thousands of images and videos daily across various platforms, continuously improving based on performance feedback (CPI and ROAS). The goal is to create a highly efficient and intelligent system for ad creative generation.

Requirements

Deep understanding of diffusion and video models, including debugging, training LoRAs, optimizing inference, and building ComfyUI workflows.
Familiarity with various generative models (Flux, SDXL, Veo, Sora, Runway, Kling, Hailuo) and their optimal use cases based on cost and quality.
Publicly available artifacts demonstrating generative AI work (e.g., fal demo, Replicate model, Civitai upload, GitHub workflow).
Instinct for creating scroll-stopping ad content that captures attention within the first 3 seconds.
Experience considering cost as a primary design constraint (e.g., $/asset, $/click, $/install) and shipping cost arbitrage strategies.
Thorough understanding of Meta Advertising Standards and TikTok branded-content rules.
Proven ability to build end-to-end pipelines, from idea to shipped product and iterative improvement based on data.

Nice To Haves

Experience with pretrained or significantly advanced VLMs.
Open-sourced a ComfyUI custom node, diffusers extension, or Replicate model with significant traction.
Experience deploying serverless inference on fal, Replicate, or Modal at scale (>10k req/day).
Shipped a video product end-to-end, covering script to final cut.
Authored public technical content on generative video model failure modes that is cited by other engineers.

Responsibilities

Extend LLM critic + performance-feedback regeneration from images-only to images + video + copy, in a closed-loop, unsupervised system generating thousands of variants per day.
Build a video pipeline that generates platform-ready 9:16 vertical ads (script, scene plan, b-roll, voiceover, captions) in under 90 seconds per iteration, at under $0.50 per iteration.
Develop a generative model router to select the optimal model (e.g., Flux Schnell vs Pro, Veo 3, Sora, Runway Gen-3/4, Kling, Hailuo) for each brief, minimizing cost per asset while maintaining a quality floor, and prove the strategy via A/B testing.
Create a VLM (Vision-Language Model) policy pre-flight checker (using models like GPT-4o vision, Gemini, or Claude vision) to vet every frame and caption against current ad policies before upload, aiming to block 100% of non-compliant creatives.
Train and deploy brand-style LoRA/DreamBooth checkpoints based on past CPI-winning ads, using serverless deployment on platforms like fal or Replicate, and A/B test against base Flux.
Develop an event-triggered video ad pipeline that can generate multiple paused video variants with platform-specific aspect ratios and locale-specific voiceovers within 10 minutes of an event.