Senior Generative Video ML Engineer, Post-Training & Fine-Tuning

Eros is building sovereign cultural AI: systems designed to make advanced intelligence culturally relevant, rights-aware, governable, and commercially useful. We are bringing together a major film and media catalog, a new AI platform, and a founding technical team, with every training and evaluation asset subject to rights, consent, provenance, territory, and permitted-use controls.

We are hiring a senior, hands-on ML engineer as a founding member of a new generative-video team. The work centers on open-weight video models, reproducible inference, controlled post-training or adaptation, and measurable improvement in character, scene, and temporal consistency.

This is not a prompt-engineering role and not a general AI-app build. We need someone who has personally trained, adapted, evaluated, and debugged modern video-generation models and can show what changed, why it changed, and how the result was measured.

What you will own

• Stand up a reproducible, versioned inference and experimentation environment for one or more open-weight video-generation models.
• Establish frozen baseline results before adaptation begins.
• Inspect data readiness, define train, validation, and sealed holdout splits, and prevent identity or scene leakage.
• Design and run license-permitted post-training experiments using the lightest justified method, such as LoRA, adapter training, supervised tuning, preference optimization, or conditioning changes.
• Build or integrate structured character and scene conditioning while preserving immutable run manifests.
• Measure identity consistency, appearance continuity, scene adherence, action fidelity, temporal stability, latency, and cost.
• Diagnose failures and compare controlled remediation strategies such as repair, retry, reroute, or rejection.
• Package code, configs, run logs, checkpoints or adapters, evaluation results, architecture notes, and complete technical handover materials.

Environment and constraints

• The work draws on a major film and media catalog, but every training and evaluation asset must pass documented rights, consent, provenance, territory, and permitted-use controls.
• All proprietary data and resulting artifacts remain inside a controlled private environment.
• No data may be copied to personal storage or external inference services.
• Training and model use must follow checkpoint-specific commercial-license, data-rights, consent, provenance, security, and territory approvals.
• Work will begin with bounded evidence phases before production-scale training runs.
• Results will be judged against frozen baselines and held-out cases. We do not accept self-selected demos as proof.

Required experience

• Deep hands-on experience with diffusion or flow-based video generation, transformer-based video models, or closely related multimodal generation systems.
• Personally executed post-training or adaptation of an open-weight image or video model using PyTorch.
• Strong GPU systems knowledge, including distributed training or inference, memory optimization, mixed precision, checkpointing, and experiment tracking.
• Built evaluation pipelines for identity, visual consistency, temporal quality, prompt or scene adherence, or production usability.
• Comfortable operating under strict data custody, reproducibility, and evidence requirements.

Strong pluses

• Direct experience with model families such as Wan, MiniMax, HunyuanVideo, CogVideoX, LTX Video, Mochi, or comparable systems.
• Character-consistency techniques, reference conditioning, identity embeddings, temporal adapters, video inpainting, or localized repair.
• Experience optimizing inference on H100 or H200-class infrastructure.
• Experience designing blinded human review alongside automated evaluation.

To apply, please address these questions

1. Describe the most relevant video-generation model you personally trained or adapted. Name the base model, method, data scale, compute, your exact contribution, and the measured before-and-after result.
2. How did you evaluate identity consistency and temporal quality on held-out examples? Include metrics, human review, and one failure your evaluation caught.
3. What is the largest multi-GPU training or inference job you personally operated, and what failed during the run?
4. Share one sanitized artifact you can walk through live, such as code, config, an experiment report, an evaluation dashboard, or an architecture document.
5. Are you able to work full time and provide reliable overlap with a distributed team?