Principal AI Engineer

PRINCIPAL AI ENGINEER

Key Responsibilities:
 Own the SMA Research Agenda: Set and drive the research roadmap for company's Specialized Model. Frame the open problems in domain distillation, decide which experiments are worth running, and own the results end to end. Your work should raise the accuracy-per-dollar frontier for the specific enterprise domains our customers care about, not for public benchmarks.
 Push Distillation and Model Compression: Design and run the teacher-student distillation programs that turn frontier models into small task-specialized models. Work on the underlying techniques (curriculum design, synthetic data generation, distillation objectives, quantization-aware training, activation compression, sparsity) and prove their value on real production workloads.
 Advance RL and Preference Optimization: Design reinforcement learning and preference optimization pipelines (RLHF, RLAIF, DPO and its successors, policy optimization) that let our agents learn from real deployment feedback. Own the reward modeling, the training loops, and the evaluation methodology that tells us whether a change actually helped.
 Own Inference Economics: Drive measurable improvements in cost-per-inference, latency, and throughput for the models we serve. This spans model-side work (distillation, quantization, speculative decoding, KV-cache design) and serving-side work (batching, scheduling, tensor parallelism, hardware selection). Cost-per-inference is a business metric here, not a benchmark.
 Bridge Research and Production: Take research that lives in papers, notebooks, or lab code and turn it into something that ships. Partner with the model runtime team on how models get served in production, and with the FDE pods on how models perform in messy customer environments. The bar is not that the paper reproduces. The bar is that the customer’s metric moved.
 Set the Technical Bar: You are the technical voice for applied ML . Run the design reviews for major model work, review experiment plans, push back on shortcuts, and make the calls on which approaches are worth investing in and which are dead ends. Your judgment should be one of the reasons a senior engineer wants to join the company.
 Grow the Team: Mentor senior AI engineers and research engineers on experiment design, evaluation rigor, and the discipline of applied research. Help hire the next wave of the ML org and set the bar in interviews. You will not carry direct reports, but the seniors on the team should be growing faster because you are here.
 Represent company Externally: Publish, present, and contribute to the applied AI research community. Speak at conferences, engage with academic collaborators, and help build company's technical reputation. Not every project will be publishable, but the ones that are should reach the venues that matter.
Required Skills & Experience:
 PhD in an AI-Adjacent Field (required): A PhD is a hard requirement for this role. Preferred research areas include machine learning, deep learning, natural language processing, large language models, reinforcement learning, knowledge distillation and model compression, efficient inference and ML systems, multimodal learning, agent-based AI, and information retrieval. Strong preference for candidates from globally reputed institutions (for example, MIT, Stanford, UC Berkeley, CMU, Princeton, Harvard, Cornell, Oxford, Cambridge, Imperial, ETH Zurich, EPFL, Toronto, Mila, NUS, IISc, or the top IITs), and for PhD work in a sub-area that directly maps to what company builds.
 Applied Research Track Record: A combined 8+ years of PhD and industry experience, or 5+ years of post-PhD applied experience, focused on shipping ML systems. Publications at top venues (NeurIPS, ICML, ICLR, ACL, EMNLP, CVPR, AAAI, or equivalent) are a strong signal, particularly first-author work. Widely used open-source contributions and citations from other researchers count for the same reason.
 Production ML at Scale: You have shipped ML systems that serve real users in production, not just research prototypes. You have owned the full arc from research idea to deployed model, and you understand what breaks in that transition. Direct experience with LLM fine-tuning, distillation, or serving at scale is expected.
 Programming and ML Frameworks: Excellent Python. Deep familiarity with PyTorch and the surrounding tooling (HuggingFace Transformers, Accelerate, DeepSpeed, vLLM, or equivalent). Comfort reading CUDA and Triton kernels, or the good judgment to know when to bring in someone who lives in them.
 Distillation and Model Efficiency Depth: Direct, shipped experience with knowledge distillation, quantization, pruning, or other model compression techniques. You have made a real model smaller, faster, or cheaper without losing the accuracy that mattered. Familiarity with inference optimization stacks (TensorRT, ONNX, vLLM, TGI) is expected.
 Reinforcement Learning Depth: Working knowledge of modern RL and preference optimization methods (PPO, DPO, GRPO, RLHF, RLAIF). You have implemented or extended these methods, not just consumed them from a library.
 Evaluation Rigor: You know how to design an evaluation that actually tells you whether a change helped. You have opinions on the failure modes of common benchmarks and on when a benchmark result does or does not translate to real-world usefulness.
 Distributed Training Systems: Comfort with distributed training at multi-node scale. You understand the tradeoffs between data parallelism, tensor parallelism, and pipeline parallelism, and you have debugged real training runs that fell over for real reasons.
 Communication and Written Depth: You can write a paper, a design doc, or an internal RFC that other researchers and engineers take seriously. You can run a technical review where the right things get debated and decided. Strong written communication is a non-negotiable for this role.

Nice to Have:
 Prior applied research or research-engineering experience at a leading AI lab or an AI-first product company.
 Contributions to widely used open-source ML projects.
 Experience partnering with academic collaborators on joint publications or standards work.
 Startup or founder-adjacent wiring, with a track record of shipping under ambiguity.

Who This Role Is For:
This is a rare seat. It sits at the intersection of applied AI research, ML systems, and production engineering for enterprise agents. It will appeal to a specific kind of engineer:
 The PhD researcher who has spent time inside a leading lab or an AI-first company and now wants their work to move a customer metric rather than a leaderboard.
 The Staff or Principal IC at a large tech company who wants a problem space with real depth and real ownership, without the overhead of a management ladder.
 The researcher who is energized by the distillation, RL, and inference-economics stack, and who thinks the interesting work in the next few years happens where those three meet.
 The engineer who is comfortable in ambiguity, comfortable making architectural calls with incomplete information, and comfortable being wrong in public when the evidence says so. This role is not for someone looking for a stable, well-defined research job with a fixed publication cadence. This company is in an active build phase, priorities shift as customers land, and the first six months will involve real tradeoffs between publishable work and shipped work. If that excites you, read on.

Cultural Fit (What We Value):
 Research Depth With Shipping Bias: You care about correctness, elegance, and the honesty of a clean ablation. You also care about whether the customer’s metric moved. These are not in tension for you.
 Ownership and Initiative: You take initiative. You do not wait to be assigned. You see a problem worth solving and you find a way to solve it, then bring others along.
 Curiosity and Learning Mindset: You stay close to the frontier of the field, not because you have to, but because you find it interesting. You read papers, you reproduce results, and you can tell what is hype and what is real.
 Collaboration Across Boundaries: You work well with product, with FDEs inside customer accounts, with backend engineers on the runtime, and with junior AI engineers. You are generous with your time and your knowledge.
 Pragmatic Impact: You are motivated by applied research that gets used. A shipped model that changes what a customer can do beats a paper that no one reads.