Research Engineer

About the Role

Deloitte is building an AI-first effort with significant investment to rebuild decision-making in the American healthcare system. This role focuses on designing, training, evaluating, and aligning models for healthcare decisioning. The work involves deep post-training techniques, not prompt engineering.

Key Responsibilities

Post-training & Alignment

• Design and execute post-training pipelines: supervised fine-tuning (SFT), preference optimization, and reinforcement learning / alignment workflows.
• Build and optimize training using techniques such as SFT, RLHF, PPO, DPO, GRPO, RLAIF, and Constitutional AI, understanding their impact on reasoning quality, safety, latency, cost, and reliability.
• Train reasoning models for healthcare decisioning using verifiable-reward RL, designing reward signals grounded in clinical guidelines, policy, criteria, and adjudicated outcomes.

Reward Modeling & Data

• Develop reward models and preference datasets to improve reasoning quality, factuality, safety, policy adherence, and task performance.
• Curate, clean, synthesize, and evaluate large-scale instruction, preference, and domain-specific datasets with rigorous filtering, deduplication, and quality control.
• Build verification and reward pipelines from proprietary clinical, claims, and operational data, and from clinical-expert labeling to create checkable reward signals.

Efficient Fine-tuning, Training & Inference Infrastructure

• Implement efficient fine-tuning strategies including LoRA, QLoRA, PEFT, and adapter-based approaches.
• Build scalable distributed training using DeepSpeed, FSDP, Megatron-LM, Ray, or equivalent.
• Optimize inference performance (latency, throughput, quantization, deployment efficiency) for production using frameworks such as vLLM, TensorRT-LLM, or TGI.

Small Language Models & Open-Weight Models

• Train and optimize open-weight models such as Llama, Qwen, Mistral, or DeepSeek.
• Build specialized small language models (SLMs) for on-premise and cloud-hybrid deployment with strong performance-per-dollar.

Evaluation, Safety & Red Teaming

• Design evaluation frameworks covering reasoning, hallucination detection, factuality, instruction following, structured outputs, and domain-specific metrics.
• Build healthcare-grade evaluation, including held-out clinical benchmarks, deployment regression gates, calibration, uncertainty, factuality against ground truth, and bias/fairness evaluation.
• Apply PHI/HIPAA-aware data handling and produce model documentation suitable for regulated clinical use.
• Perform red teaming and adversarial testing to identify alignment failures, unsafe behaviors, jailbreak vulnerabilities, and regression risks.
• Collaborate with agentic and application teams to improve tool use, grounding, and long-horizon reasoning.

Requirements & Tech Stack

• Post-training & Alignment: SFT, RLHF, PPO, DPO, GRPO, RLAIF, Constitutional AI, preference optimization.
• Reward Modeling & Data: Reward modeling, preference datasets, large-scale datasets, data curation, filtering, deduplication, quality control.
• Efficient Training & Inference: LoRA, QLoRA, PEFT, distributed training (DeepSpeed, FSDP, Megatron-LM, Ray), inference optimization (vLLM, TensorRT-LLM, TGI).
• Models: Small Language Models (SLMs), open-weight models (Llama, Qwen, Mistral, DeepSeek).
• Evaluation & Safety: Evaluation frameworks, hallucination detection, factuality, PHI/HIPAA-aware data handling, red teaming, adversarial testing.
• Programming: Python engineering skills, production-grade software practices.
• Deep Learning: PyTorch, modern deep-learning tooling, transformer architectures, tokenization, attention mechanisms, decoding strategies, model scaling trade-offs.
• Other: Verifiable-reward (RLVR) workflows.

Qualifications

• Education: Bachelor's degree in Computer Science, Machine Learning, Artificial Intelligence, Applied Mathematics, Computational Linguistics, or a related field.
• Experience:
• Demonstrated depth training and post-training large transformer-based language models in production or research.
• Genuine depth including SFT and at least one preference-optimization or RL method, evidenced by shipped models, releases, or research.
• Hands-on experience with reasoning-model training and/or verifiable-reward (RLVR) workflows.
• Strong understanding of modern post-training techniques: SFT, RLHF, PPO, DPO, GRPO, RLAIF, and preference optimization workflows.
• Experience with open-weight foundation models such as Llama, Qwen, Mistral, DeepSeek, or equivalent architectures.
• Strong expertise in PyTorch and modern deep-learning tooling; experience with distributed training frameworks such as DeepSpeed, FSDP, Megatron-LM, or Ray.
• Experience implementing efficient fine-tuning techniques such as LoRA, QLoRA, PEFT, and quantization-aware workflows.
• Deep understanding of transformer architectures, tokenization, attention mechanisms, decoding strategies, and model scaling trade-offs.
• Strong grasp of LLM evaluation methodologies, benchmarking, reward modeling, and alignment trade-offs.
• Experience with large-scale and synthetic datasets, filtering, deduplication, and quality-control pipelines.
• Ability to work through ambiguous, highly complex technical problems in fast-moving environments.
• Familiarity with inference optimization frameworks such as vLLM, TensorRT-LLM, TGI, or Ollama.
• Experience with multimodal models, speech models, or domain-specific foundation models.
• Experience using large-scale GPU clusters and distributed compute.
• Contributions to open-source AI projects, research publications, benchmark development, or model releases.
• Familiarity with safety, governance, and responsible-AI practices; experience in regulated or high-stakes industries such as healthcare, finance, insurance, or public sector.
• Travel: Ability to travel 0-50%, on average.
• Immigration Sponsorship: Limited immigration sponsorship may be available.

Compensation

• Wage Range: $110,700 - $379,200