Senior AI/ML Developer

<p>→Own and drive the full RLHF pipeline: data collection, reward model training, and RL fine-tuning using PPO, DPO, GRPO, and RLAIF</p><p>→Design and run Supervised Fine-Tuning (SFT) pipelines on open-weight models (LLaMA, Mistral, Qwen) as the foundation for RLHF</p><p>→Build and train reward models that accurately capture human preferences from annotation data</p><p>→Design human feedback collection pipelines: labeling rubrics, annotator calibration, and preference dataset curation</p><p>→Implement Constitutional AI and RLAIF techniques to reduce reliance on costly human annotation</p><p>→Red team models post-training — probing for jailbreaks, regressions, unsafe outputs, and alignment failures</p><p>→Design and maintain evaluation benchmarks to measure alignment, safety, and capability before and after RL training</p><p>→Optimize inference pipelines and runtimes (llama.cpp, vLLM, TensorRT) to serve aligned models efficiently at scale</p><p>→Implement quantization strategies (INT4/INT8/FP8, LoRA, QLoRA) to deploy fine-tuned models on target hardware</p><p>→Write and tune low-level C/C++ and Rust code for inference performance where Python cannot reach</p><p>→Diagnose and resolve training instabilities, reward hacking, and production inference bugs under pressure</p><p>→Stay at the frontier — read alignment and RL papers weekly and translate findings into working experiments</p><div><div><b>Core Requirements and Technical Skills</b></div><div>→Hands-on experience implementing RLHF end-to-end — not just using libraries, but understanding the mechanics</div><div>→Deep familiarity with policy gradient methods: PPO stability, KL divergence constraints, reward shaping</div><div>→Experience with Direct Preference Optimization (DPO) and its variants as an RLHF alternative</div><div>→Understanding of reward hacking, Goodhart’s Law, and mitigation strategies in RL training</div><div>→Familiarity with RLAIF (RL from AI Feedback) and Constitutional AI approaches</div><div>→Ability to design preference datasets and annotation rubrics that produce high-quality reward signal</div><div>→Experience diagnosing training instabilities: reward collapse, mode collapse, KL divergence blowup</div><div>→Python as the primary language for all training, fine-tuning, and evaluation pipelines</div><div>→Strong mathematical foundation: RL theory, probability, linear algebra, optimization — deep enough to derive loss functions and debug training dynamics</div><div>→C and C++ for systems-level inference work, runtime contributions, and performance-critical paths</div><div>→Rust experience with ML tooling.</div><div>→Familiarity with transformer architecture, attention, tokenization, and how post-training interacts with pretraining</div><div>→Experience with distributed training frameworks for large-scale fine-tuning</div><div>→Experience with vector databases such as FAISS or Milvus</div><div>→Familiarity with retrieval-augmented generation (RAG) pipelines</div><div>→Experience integrating LLMs with external tools, APIs, and agent-based systems</div><div>→Exposure to Rapid Application Development (RAD) approaches for building and iterating AI solutions efficiently</div></div><div><br /></div>