AI Engineer, Gopuff

• PhD in Computer Science, Machine Learning, or equivalent research experience with significant contributions to AI/ML literature
• 7+ years of building and shipping large‑scale ML systems with significant ownership; proven impact in production LLM or RL‑driven products
• Mastery of advanced fine-tuning techniques including LoRA/QLoRA, adapter methods, and parameter-efficient transfer learning
• Research experience with agentic AI frameworks, multi-agent systems, and declarative programming approaches (DSPy, LangChain ecosystem)
• Strong systems engineering capabilities with PyTorch, distributed training, and cloud-native ML infrastructure
• Track record of publications in top-tier venues (NeurIPS, ICML, ICLR, AAAI) or equivalent industry impact
• Deep expertise in transformer architectures, SFT, and RLHF; hands‑on leadership with RLVR and verifiable reward design
• Mastery of policy optimization (DPO/PPO/GRPO/GSPO) and the ability to extend/regularize policies under safety, latency, and cost constraints
• Strong grounding in offline evaluation, counterfactual estimators, and safe online ramp strategies
• Systems fluency: PyTorch, distributed training, low‑latency serving, observability, and cloud‑native ML infra
• Demonstrated leadership across cross‑functional teams, with clear communication and mentoring track record
• Commitment to responsible AI: privacy, safety, and alignment principles embedded end‑to‑end

Desirable

• As an AI Engineer III, you will be the technical lead across context engineering, RLHF/RLVR and low‑latency serving
• You’ll define the architecture, standards, and evaluation strategy that connect research to real‑world lift
• You’ll mentor colleagues, influence cross‑functional roadmaps, and ship systems that deliver measurable improvements to core customer and business outcomes—without disclosing competitive intelligence
• Set the strategy for context engineering to maximize precision/recall of key order metrics across sessions, households, locales, and time
• Architect multi‑modal context integration (temporal, spatial, behavioral) and real‑time grounding with dynamic constraint satisfaction
• Establish retrieval freshness, geo/time‑aware constraints, and memory policies; formalize context schemas and data contracts
• Champion declarative prompt/program compilation (e.g., DSPy) for systematic, testable LLM behavior
• Design multi‑agent orchestration patterns (e.g., graph‑based agents via LangChain/LangGraph, CrewAI, AutoGen, LlamaIndex) that yield robust emergent reasoning
• Lead supervised reasoning-centered fine‑tuning with rigorous data curation, synthetic data generation, and QA; institute golden sets and rubric/pairwise evals
• Own the reasoning architecture and evaluation strategy—planning, tool selection, reflection, and uncertainty-aware decision-making—to deliver robust, low-latency, grounded outcomes at scale
• Drive parameter‑efficient adaptation strategies (LoRA/QLoRA and text-to-LoRA) with clear criteria for when to specialize vs. generalize
• Architect RLHF and RLVR pipelines; build preference data loops, scalable oversight, and guardrails
• Own policy optimization strategy: expert use of DPO/PPO/GRPO/GSPO and advancement beyond them (constrained optimization, regularized objectives, KL‑control) with formal safety considerations
• Ensure robust offline‑to‑online correlation via counterfactual/IPS/DR estimators and stress tests across traffic segments
• Establish interpretability, controllability, and alignment verification practices for agentic systems
• Develop safeguards against reward hacking and unsafe exploration; enforce distributional robustness and content policy compliance
• Advance privacy‑preserving methods (data minimization, federated/on‑device learning where appropriate) with privacy‑by‑design
• Architect low‑latency, cost‑efficient inference (quantization, caching, batching, streaming) with resilient fallbacks and red‑teaming
• Build eval frameworks that tightly couple offline metrics with online performance and safety criteria; define promotion gates
• Use relevant APIs to perform high‑fidelity data augmentation that strengthens grounding, disambiguation, and availability‑aware suggestions
• Partner closely with Engineering and Data Science to design experiments, define success criteria, and iterate quickly from signal to lift
• Translate ambiguous product goals into crisp technical milestones; maintain clear documentation, incident response, and learning playbooks
• Mentor colleagues; raise the bar on design quality, reproducibility, and ethical rigor
• Tooling and Stack:
• Fine‑tuning: Unsloth for rapid prototyping; TRL for RLHF/RLVR workflows and policy optimization
• Retrieval, evaluation, and orchestration: pragmatic use of graph‑based agent frameworks and vector retrieval systems as appropriate