AI Engineer, Agentic Ad Creative (Multimodal)

About the Role

Every install we buy starts as a creative someone (or something) made. We're rebuilding that creative pipeline so that *something* is an agent — one that generates thousands of images and videos a day across Meta, Google, TikTok, Pinterest, Reddit, and Snap; watches CPI and ROAS come back; and gets sharper every cycle.

We're hiring the engineer who owns this agent end-to-end.

What You'll Do (representative projects)

• CPI-feedback creative critic loop — Extend our LLM critic + performance-feedback regeneration from images-only to images + video + copy, all closed-loop, all running unsupervised at thousands of variants per day.
• Video pipeline — Build from a topic + product brief to a platform-ready 9:16 vertical ad — script, scene plan, b-roll, voiceover (ElevenLabs / Hedra), captions — in under 90 seconds per iteration, under $0.50 per iteration.
• Generative model router — Pick the right model per brief (Flux Schnell vs Pro, Veo 3, Sora, Runway Gen-3/4, Kling, Hailuo), minimizing $/asset under a quality floor. Prove your routing strategy in production via A/B.
• Policy pre-flight checker — Build a VLM (GPT-4o vision / Gemini / Claude vision) that vets every frame and caption against current Meta / Google / TikTok ad policy before upload. Block 100% of would-be-rejected creatives at the source.
• Brand-style LoRA / DreamBooth checkpoint — Train and ship on past CPI-winning ads. Deploy serverless on fal / Replicate. A/B against base Flux on real spend.
• Event-triggered ad pipeline (video) — Earthquake hits at 14:02 by 14:12 you have 50 paused video variants in Meta and TikTok, each with platform-specific aspect ratios and locale-specific voiceover.

You May Be a Good Fit If

• You understand diffusion and video models viscerally — you've debugged them, trained a LoRA, optimized inference, built a ComfyUI workflow you'd defend in public.
• You can name your daily model stack — Flux vs SDXL vs Veo vs Sora vs Runway vs Kling vs Hailuo — and explain when each wins, where each breaks, and what you'd reach for at $0.01/asset vs $0.50/asset.
• You have public artifacts — a fal demo, a Replicate model, a Civitai upload with >100 downloads, a viral X thread of your generations, a ComfyUI workflow on GitHub. Portfolio matters more than resume.
• You have scroll-stopping instinct — you know what makes someone not swipe past an ad in the first 3 seconds.
• You think about cost as a first-class design constraint — $/asset, $/click, $/install. Cost arbitrage between model tiers is something you've actually shipped.
• You've read Meta Advertising Standards / TikTok branded-content rules end-to-end — policy is a pre-deploy gate, not a post-mortem.
• You're a builder — Idea working pipeline first 100 ads shipped CPI data back tighter loop. Not "I'll spec it for Q3."

Strong Candidates May Also Have

• Pretrained or significantly advanced a VLM (not just SFT'd or LoRA'd one).
• Open-sourced a ComfyUI custom node, a diffusers extension, or a Replicate model with real traction.
• Deployed serverless inference on fal / Replicate / Modal at scale (>10k req/day).
• Shipped a video product end-to-end: script scene plan b-roll voiceover captions final cut.
• Authored a public take on generative video model failure modes that other engineers cite.

Not the Right Fit If

• You've only ever prompted hosted models — you've never trained a LoRA, written a ComfyUI workflow, or deployed a custom checkpoint.
• Your creative experience stops at Photoshop or Canva templates. You can't tell us why a Flux generation has bad hands or why Veo 3 over-saturates skin.
• You think "AI creative" means asking GPT to write ad copy.
• You're an academic researcher chasing first-author papers. We ship.
• You're a brand-side creative director without engineering depth. We need someone who debugged a CUDA OOM at 2am because the generator was holding up the next batch.

How We'll Evaluate You

Take-home: build a small agentic creative pipeline end-to-end — your call: image-only or short video, hosted or self-served, model of your choice. Walk us through what you built and every model / cost / quality decision behind it. Conversations with the team follow. Portfolio matters more than resume — if you can't show a public artifact, please come back when you can.

What You Get

• Competitive Bay Area comp + meaningful equity.
• Unlimited generative model budget — fal, Veo, Sora, Runway, ElevenLabs, OpenAI image, Anthropic vision, Hedra, Kling. We expect you to burn through it.

About the Team

You'd join the team building NewsBreak's agentic growth platform. We already ship a CPI-feedback creative critic loop, event-driven creative pipelines, and multi-platform ad-upload tooling for Meta, Google, and TikTok. Your job: take the creative side from "thousands of images per day" to "video too, smarter every cycle, cheaper every month."

If you read this and thought "finally, a JD that's written for me" — we want to talk.

Benefits

We offer a competitive benefits package:

• Health, dental, and vision care for you and your family (100% coverage for employee)
• Top-tier 401(K) plan with company matching
• Paid time off and paid holidays
• FSA, HSA and commuter benefits programs
• Team activity budget