1 year ago
Responsibilities
- Fine-tune and evaluate open-source LLMs for enterprise use cases involving structured and unstructured data.
- Build RAG pipelines using agent orchestration and retrieval frameworks integrated with vector databases and knowledge graphs.
- Train and customize agent architectures with multi-step reasoning, memory, tool use, and multi-agent communication.
- Create reinforcement-learning pipelines using RLHF, DPO, and PPO to optimize agent behavior.
- Develop scalable evaluation harnesses with synthetic evaluations, trace capture, explainability, observability, drift detection, and error classification.
- Optimize inference latency, token costs, quantized and multi-GPU execution, and GPU resource utilization across cloud and on-premise environments.
- Own model performance end-to-end from data curation and fine-tuning through evaluation and production deployment.
- Collaborate with product and UX teams to shape AI capabilities.
Requirements
- Deep experience fine-tuning open-source LLMs with HuggingFace Transformers, DeepSpeed, vLLM, FSDP, and LoRA or QLoRA.
- Experience working with base and instruction-tuned models and SFT, RLHF, and DPO pipelines.
- Ability to build custom training datasets, filters, evaluation splits, and model evaluation systems.
- Understanding of batch size, token windows, optimizers, numerical precision, and quantization tradeoffs.
- Experience building enterprise RAG pipelines with real-time or contextual data, vector databases, and structured and unstructured sources.
- Familiarity with agent frameworks, multi-step reasoning, memory, tools, self-correction, multi-agent communication, and agent operations logging.
- Strong background in token-cost optimization, chunking, reranking, compression, retrieval latency tuning, quantized inference, and multi-GPU systems.
- Preferred experience with Neo4j, Puppygraph, RDF, OWL, semantic modeling systems, and the listed evaluation, compute, storage, and orchestration technologies.
- Strong interest in agent-based architectures, enterprise complexity, explainability, safety, and continuous improvement.
Benefits
- Full-time founding-team role in the San Francisco Bay Area.
- Meaningful equity at the founding tier.
- Opportunity to productionize AI infrastructure used by hundreds of users and support high-impact enterprise workflows.
Tech Stack
Categories
About Fabrion
Fabrion builds enterprise AI infrastructure centered on agent-based systems, knowledge graphs, and multi-tenant governance for mission-critical workflows. The privately held company operates an Agentic AI Lab to train and evaluate specialized models, offering software and platform services to large organizations; founded in 2025 and headquartered in San Francisco.