AI Engineer (Ad Performance and Video Understanding)

The problem you'll solve

We spend heavily on video ads that find people hurt in accidents and connect them with law firms. The only number that matters is cost per signed case, and a signed case often arrives weeks after the click.

We've built an internal app that joins ad delivery from Meta, Google and TikTok with the signed cases in our CRM. Its AI already watches every ad, tags its hooks, scenes and structure, reviews editors' cuts, and proposes the next brief. A person always makes the decision.

Two things are hard. The first is understanding video. Today the AI describes what is in an ad. We need it to understand why an ad was cut the way it was, which seconds lose people, which creative choices lead to signed cases rather than cheap leads, and what an editor should change next.

The second is judging ads the way we do. Most advertisers chase cheaper leads. We don't. We'll gladly pay several times more for a lead that signs than for one that never does. Meta and Google push budget toward whatever we tell them counts, so the AI has to understand how campaigns, ad sets, learning and fatigue work, and judge every ad on calls and signed cases, not on cost per lead.

You'll make the AI the second brain of our creative strategists, video editors and media buyers. It won't replace them. It will help them see what is working and act on it the same day.

What you'll build

1. Judge every ad on what it really brings in. Tie spend in Meta, Google and TikTok to calls, qualified leads and signed cases. Show which campaigns, ad sets and ads are failing, and why.

2. Moves for media buyers. Early warnings on ads that are fatiguing or bringing leads that won't sign, and where to shift budget, while respecting how each platform's learning reacts to changes.

3. Video understanding at scale. A pipeline that breaks every ad into shots, scenes, speech, on-screen text, pacing and edit choices, and stores them as features we can model. Hundreds of ads today, thousands later.

4. Why a video works. Models that link those features to watch time by scene and to signed cases, not just clicks. They must say how sure they are, because signed cases are few and slow.

5. Editing guidance. Feedback an editor can act on, tied to exact seconds: cut here, move the hook, this scene loses people compared with our typical ad. Later, the app produces rough cuts, hook swaps and variants itself.

6. Angles and briefs. Creative angles, hooks and scripts drawn from what has actually signed cases, so strategists start each brief from evidence.

7. Proof the AI is right. Evaluation sets, saved predictions checked later against real results, and agreement checks with our team's own labels.

You'll work in the existing app (TypeScript, Next.js, Postgres) beside our full-stack engineer, who owns the data syncs and the interface. You own the models, the video pipeline and how we measure them. Python services are fine where they're the right tool.

What you must have

• Hands-on paid ads experience. You've worked inside Meta Ads or Google Ads data: campaign, ad set and ad structure, attribution windows, learning phases, fatigue, and why campaigns stop working. You know media buying well enough to argue with a buyer.

• Optimized for what happens after the lead. You've built or run a system that judged ads or campaigns on calls, qualified leads, sales or retention, not on clicks or cost per lead. Sending CRM or offline conversions back to an ad platform is a strong sign.

• Video or audio ML in production. You've personally built and shipped a system that processes real video or audio: shot or scene detection, speech-to-text, frame or clip embeddings, or multimodal models over video. A demo or a course project doesn't count.

• Honest numbers on small, slow data. You've linked content or campaign features to a business result and can explain how you avoided fooling yourself when the outcomes were few or delayed.

• Evaluation you can show. You've built test sets and measured a model's accuracy, then used the numbers to decide what shipped.

• Multimodal LLMs used well. Structured outputs, frame sampling, cost and latency trade-offs, and knowing when a smaller trained model beats a large prompted one.

• Strong Python and solid engineering. Data pipelines, queues or batch jobs, Postgres, and enough TypeScript to work in a Next.js codebase.

• 5+ years of software or ML engineering, with at least 2 on ML or marketing-data systems used by real people.

Nice to have

• Lead generation where a lead's value shows up weeks later: legal, insurance, home services, education

• TikTok or YouTube ads data

• You've worked closely with video editors or edited video yourself

• ffmpeg, video rendering, or generating video variants

• Fine-tuning or training vision or audio models

• pgvector or other vector search in production

What disqualifies you

• No hands-on work with ad platform or marketing data: video or ML experience alone isn't enough

• No production work with video or audio: only chatbots, RAG over documents, or generic LLM apps

• Has only judged ads on clicks, click-through rate or cost per lead

• Can't name a metric you measured and the number it reached

• A résumé that repeats this post's wording, or claims you can't explain live on camera

• Unwilling to walk through your code or past work on a video call

Role details

Type

Full-time, long-term contract

Location

Remote: United States, India or Pakistan

Hours

At least 4 hours a day overlapping US Central time

Team

Our full-stack engineer, our creative strategists, video editors and media buyers

Payment

Biweekly or monthly, by Wise, Payoneer or wire

Agreements

NDA and full IP assignment

How hiring works

1. Application with the qualifier and written questions on our website

2. 45-minute video call, cameras on, walking through your past work

3. Technical call with our engineer on a real problem from our app

4. A short paid work sample on real ads

5. Offer