Opportunity brief
About the Company: Horrazon AI is building Pogee, their own LLM, and an agent layer to ingest live data from various social platforms, transforming it into actionable decisions for founders. They are pre-revenue, venture-track, and backed by Nvidia Inception, AWS Startup, Zoho Startup, and others. The team is small, high-leverage, and focused on quickly shipping features that impact real customers. Responsibilities: Training & Fine-Tuning: Fine-tune open-source LLMs using techniques like LoRA/QLoRA, full fine-tuning, and preference-optimization (RLHF/DPO-style) for product use cases. Model Fundamentals: Apply a deep understanding of transformer architecture, tokenization, pretraining objectives, and training dynamics. Agentic Systems: Design and build the reasoning and decision layer for agents, including multi-step planning, tool use, and pipelines for data-driven recommendations. Evaluation: Develop rigorous evaluations, conduct controlled experiments, and make evidence-backed decisions for deployment. Data Engineering for Training: Work with Python/Go/PostgreSQL to prepare and structure high-quality datasets for training and fine-tuning from multi-platform social and ads data. Efficiency: Manage inference cost, latency, and quantization decisions, with a focus on resource optimization. Requirements: Pursuing or recently completed a degree in CS, AI/ML, Data Science, or equivalent, with a strong emphasis on project history. Demonstrated, hands-on experience training or fine-tuning LLMs using PyTorch, Hugging Face Transformers/PEFT, and at least one real pretraining or fine-tuning project. Deep working knowledge of transformer internals: attention, embeddings, loss functions, tokenization. Real exposure to agentic AI, including building or experimenting with tool-calling, multi-step reasoning, or orchestration frameworks (LangChain, LlamaIndex, or custom stacks). Proficiency in Python; comfortable with SQL and REST/OAuth2 API integration. Git-fluent, writes clean and documented code, and communicates precisely in an async, remote-first environment. Strongly Preferred Multi-GPU or distributed training experience, or hands-on model quantization work. Open-source contributions to ML/LLM projects, competitive ML rankings (Kaggle, etc.), or published research. Familiarity with Go, PostgreSQL, or social platform APIs (Meta, LinkedIn, YouTube, X).