Harness Engineer

YAM is a production operating system for film and TV — the relay that moves a project from script to delivery without things falling between the cracks. Built by people who've actually shipped productions for the better part of two decades.

We're a small team at Gengis.ai, working alongside Refinery Media and X3D Studio.

Harness Engineer

YAM's hard part isn't calling a model — it's everything around it. The control loop, the evals, the routing, the fine-tuning. That work is the company. This role builds it.

What you'll do

• Build the multi-step agent loop that turns creative input into production-ready output, with correction passes between steps
• Build the evals — define what "good" means for subjective creative output, and make quality a number we can move
• Design the model gateway: route between in-house and frontier models per task, with cost and provenance tracked
• Run the post-training loop (SFT, RL with rubric rewards) on the in-house components

You'll thrive here if

• You've built agent loops or eval harnesses that ran in production — something that touched real users or real spend
• You're fluent in LLM/VLM plumbing: tool use, structured output, retries, evals, cost/latency tradeoffs
• You can read a paper, ignore 80% of it, and ship the 20% that matters this week
• You default to small models + good harness over big models + hope

Bonus

• Fine-tuned open-source models (SFT, DPO, RL with rubric rewards)
• Worked on creative-tool evals where "good" is subjective
• Film, TV, or visual-production exposure

Stack — Modern, decisions still open. Real input on what we build with.

Location — Singapore preferred. SEA-remote considered for the right person.

Comp — Full-time. Competitive cash for an early-stage company plus meaningful equity. Specifics in the first conversation.

Interview — One paid working session on a real YAM problem. No whiteboards, no take-homes you'll never hear back on.

Apply — Email [email protected] with "Harness Engineer" in the subject and one paragraph on the most interesting harness or eval you've built — and what you learned was wrong about it.